153 108 2MB
English Pages 228 [222] Year 2006
Springer Series in Statistics Advisors: P. Bickel, P. Diggle, S. Fienberg, U. Gather, I. Olkin, S. Zeger
Springer Series in Statistics Alho/Spencer: Statistical Demography and Forecasting. Andersen/Borgan/Gill/Keiding: Statistical Models Based on Counting Processes. Atkinson/Riani: Robust Diagnostic Regression Analysis. Atkinson/Riani/Cerioli: Exploring Multivariate Data with the Forward Search. Berger: Statistical Decision Theory and Bayesian Analysis, 2nd edition. Borg/Groenen: Modern Multidimensional Scaling: Theory and Applications, 2nd edition. Brockwell/Davis: Time Series: Theory and Methods, 2nd edition. Bucklew: Introduction to Rare Event Simulation. Cappé/Moulines/Rydén: Inference in Hidden Markov Models. Chan/Tong: Chaos: A Statistical Perspective. Chen/Shao/Ibrahim: Monte Carlo Methods in Bayesian Computation. Coles: An Introduction to Statistical Modeling of Extreme Values. David/Edwards: Annotated Readings in the History of Statistics. Devroye/Lugosi: Combinatorial Methods in Density Estimation. Efromovich: Nonparametric Curve Estimation: Methods, Theory, and Applications. Eggermont/LaRiccia: Maximum Penalized Likelihood Estimation, Volume I: Density Estimation. Fahrmeir/Tutz: Multivariate Statistical Modelling Based on Generalized Linear Models, 2nd edition. Fan/Yao: Nonlinear Time Series: Nonparametric and Parametric Methods. Farebrother: Fitting Linear Relationships: A History of the Calculus of Observations 1750-1900. Federer: Statistical Design and Analysis for Intercropping Experiments, Volume I: Two Crops. Federer: Statistical Design and Analysis for Intercropping Experiments, Volume II: Three or More Crops. Ferraty/Vieu: Nonparametric Functional Data Analysis: Models, Theory, Applications, and Implementation Ghosh/Ramamoorthi: Bayesian Nonparametrics. Glaz/Naus/Wallenstein: Scan Statistics. Good: Permutation Tests: Parametric and Bootstrap Tests of Hypotheses, 3rd edition. Gouriéroux: ARCH Models and Financial Applications. Gu: Smoothing Spline ANOVA Models. Györfi/Kohler/Krzyz• ak/Walk: A Distribution-Free Theory of Nonparametric Regression. Haberman: Advanced Statistics, Volume I: Description of Populations. Hall: The Bootstrap and Edgeworth Expansion. Härdle: Smoothing Techniques: With Implementation in S. Harrell: Regression Modeling Strategies: With Applications to Linear Models, Logistic Regression, and Survival Analysis. Hart: Nonparametric Smoothing and Lack-of-Fit Tests. Hastie/Tibshirani/Friedman: The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Hedayat/Sloane/Stufken: Orthogonal Arrays: Theory and Applications. Heyde: Quasi-Likelihood and its Application: A General Approach to Optimal Parameter Estimation. (continued after index)
Yves Tille´
Sampling Algorithms
Yves Tillé Institut de Statistique, Université de Neuchâtel Espace de l’Europe 4, Case postale 805 2002 Neuchâtel, Switzerland [email protected]
Library of Congress Control Number: 2005937126 ISBN-10: 0-387-30814-8 ISBN-13: 978-0387-30814-2 © 2006 Springer Science+Business Media, Inc. All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, Inc., 233 Springer Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. Printed in the United States of America. 9 8 7 6 5 4 3 2 1 springer.com
(MVY)
Preface
This book is based upon courses on sampling algorithms. After having used scattered notes for several years, I have decided to completely rewrite the material in a consistent way. The books of Brewer and Hanif (1983) and H´ ajek (1981) have been my works of reference. Brewer and Hanif (1983) have drawn up an exhaustive list of sampling methods with unequal probabilities, which was probably a very tedious work. The posthumous book of H´ ajek (1981) contains an attempt at writing a general theory for conditional Poisson sampling. Since the publication of these books, things have been improving. New techniques of sampling have been proposed, to such an extent that it is difficult to have a general idea of the interest of each of them. I do not claim to give an exhaustive list of these new methods. Above all, I would like to propose a general framework in which it will be easier to compare existing methods. Furthermore, forty-six algorithms are precisely described, which allows the reader to easily implement the described methods. This book is an opportunity to present a synthesis of my research and to develop my convictions on the question of sampling. At present, with the splitting method, it is possible to construct an infinite amount of new sampling methods with unequal probabilities. I am, however, convinced that conditional Poisson sampling is probably the best solution to the problem of sampling with unequal probabilities, although one can object that other procedures provide very similar results. Another conviction is that the joint inclusion probabilities are not used for anything. I also advocate for the use of the cube method that allows selecting balanced samples. I would also like to apologize for all the techniques that are not cited in this book. For example, I do not mention all the methods called “order sampling” because the methods for coordinating samples are not examined in this book. They could be the topic of another publication. This material is aimed at experienced statisticians who are familiar with the theory of survey sampling, to Ph.D. students who want to improve their knowledge in the theory of sampling and to practitioners who want to use or implement modern sampling procedures. The R package “sampling” available
VI
Preface
on the Web site of the Comprehensive R Archive Network (CRAN) contains an implementation of most of the described algorithms. I refer the reader to the books of Mittelhammer (1996) and Shao (2003) for questions of inferential statistics, and to the book of S¨ arndal et al. (1992) for general questions related to the theory of sampling. Finally, I would like to thank Jean-Claude Deville who taught me a lot on ´ the topic of sampling when we worked together at the Ecole Nationale de la Statistique et de l’Analyse de l’Information in Rennes from 1996 to 2000. I thank Yves-Alain Gerber, who has produced most of the figures of this book. I am also grateful to C´edric B´eguin, Ken Brewer, Lionel Qualit´e, and PaulAndr´e Salamin for their constructive comments on a previous version of this book. I am particularly indebted to Lennart Bondesson for his critical reading of the manuscript that allowed me to improve this book considerably and to Leon Jang for correction of the proofs. Neuchˆ atel, October 2005 Yves Till´e
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V 1
Introduction and Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Representativeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 The Origin of Sampling Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.1 Sampling with Unequal Probabilities . . . . . . . . . . . . . . . . . 1.3.2 Conditional Poisson Sampling . . . . . . . . . . . . . . . . . . . . . . . 1.3.3 The Splitting Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.4 Balanced Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 Scope of Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5 Aim of This Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.6 Outline of This Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1 1 1 2 2 2 3 3 4 4 5
2
Population, Sample, and Sampling Design . . . . . . . . . . . . . . . . . 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Population and Variable of Interest . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 Sample Without Replacement . . . . . . . . . . . . . . . . . . . . . . . 2.3.2 Sample With Replacement . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5 Convex Hull, Interior, and Subspaces Spanned by a Support . . 2.6 Sampling Design and Random Sample . . . . . . . . . . . . . . . . . . . . . 2.7 Reduction of a Sampling Design With Replacement . . . . . . . . . . 2.8 Expectation and Variance of a Random Sample . . . . . . . . . . . . . 2.9 Inclusion Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.10 Computation of the Inclusion Probabilities . . . . . . . . . . . . . . . . . 2.11 Characteristic Function of a Sampling Design . . . . . . . . . . . . . . . 2.12 Conditioning a Sampling Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.13 Observed Data and Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.14 Statistic and Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7 7 7 8 8 9 9 12 14 14 15 17 18 19 20 20 21
VIII
Contents
2.15 Sufficient Statistic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.16 The Hansen-Hurwitz (HH) Estimator . . . . . . . . . . . . . . . . . . . . . . 2.16.1 Estimation of a Total . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.16.2 Variance of the Hansen-Hurwitz Estimator . . . . . . . . . . . 2.16.3 Variance Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.17 The Horvitz-Thompson (HT) Estimator . . . . . . . . . . . . . . . . . . . . 2.17.1 Estimation of a Total . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.17.2 Variance of the Horvitz-Thompson Estimator . . . . . . . . . 2.17.3 Variance Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.18 More on Estimation in Sampling With Replacement . . . . . . . . .
22 26 26 26 27 28 28 28 28 29
3
Sampling Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Sampling Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Enumerative Selection of the Sample . . . . . . . . . . . . . . . . . . . . . . . 3.4 Martingale Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Sequential Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6 Draw by Draw Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7 Eliminatory Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.8 Rejective Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31 31 31 32 32 33 35 37 38
4
Simple Random Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Definition of Simple Random Sampling . . . . . . . . . . . . . . . . . . . . . 4.3 Bernoulli Sampling (BERN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.1 Sampling Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.3 Sequential Sampling Procedure for BERN . . . . . . . . . . . . 4.4 Simple Random Sampling Without Replacement (SRSWOR) . 4.4.1 Sampling Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.3 Draw by Draw Procedure for SRSWOR . . . . . . . . . . . . . . 4.4.4 Sequential Procedure for SRSWOR: The SelectionRejection Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.5 Sample Reservoir Method . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.6 Random Sorting Method for SRSWOR . . . . . . . . . . . . . . . 4.5 Bernoulli Sampling With Replacement (BERNWR) . . . . . . . . . . 4.5.1 Sampling Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.3 Sequential Procedure for BERNWR . . . . . . . . . . . . . . . . . 4.6 Simple Random Sampling With Replacement (SRSWR) . . . . . . 4.6.1 Sampling Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.6.2 Distribution of n[r(S)] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.6.3 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.6.4 Draw by Draw Procedure for SRSWR . . . . . . . . . . . . . . . .
41 41 41 43 43 44 44 45 45 46 47 48 48 50 51 51 52 53 53 53 55 57 60
Contents
IX
4.6.5 Sequential Procedure for SRSWR . . . . . . . . . . . . . . . . . . . 61 4.7 Links Between the Simple Sampling Designs . . . . . . . . . . . . . . . . 61 5
6
Unequal Probability Exponential Designs . . . . . . . . . . . . . . . . . 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 General Exponential Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.1 Minimum Kullback-Leibler Divergence . . . . . . . . . . . . . . . 5.2.2 Exponential Designs (EXP) . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Poisson Sampling Design With Replacement (POISSWR) . . . . 5.3.1 Sampling Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.3 Sequential Procedure for POISSWR . . . . . . . . . . . . . . . . . 5.4 Multinomial Design (MULTI) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.1 Sampling Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.3 Sequential Procedure for Multinomial Design . . . . . . . . . 5.4.4 Draw by Draw Procedure for Multinomial Design . . . . . . 5.5 Poisson Sampling Without Replacement (POISSWOR) . . . . . . . 5.5.1 Sampling Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.2 Distribution of n(S) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.3 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.4 Sequential Procedure for POISSWOR . . . . . . . . . . . . . . . . 5.6 Conditional Poisson Sampling (CPS) . . . . . . . . . . . . . . . . . . . . . . . 5.6.1 Sampling Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.6.2 Inclusion Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.6.3 Computation of λ from Predetermined Inclusion Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.6.4 Joint Inclusion Probabilities . . . . . . . . . . . . . . . . . . . . . . . . 5.6.5 Joint Inclusion Probabilities: Deville’s Technique . . . . . . 5.6.6 Computation of α(λ, Sn ) . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.6.7 Poisson Rejective Procedure for CPS . . . . . . . . . . . . . . . . . 5.6.8 Rejective Procedure with Multinomial Design for CPS . 5.6.9 Sequential Procedure for CPS . . . . . . . . . . . . . . . . . . . . . . . 5.6.10 Alternate Method for Computing π from λ . . . . . . . . . . . 5.6.11 Draw by Draw Procedure for CPS . . . . . . . . . . . . . . . . . . . 5.7 Links Between the Exponential Designs . . . . . . . . . . . . . . . . . . . . 5.8 Links Between Exponential Designs and Simple Designs . . . . . . 5.9 Exponential Procedures in Brewer and Hanif . . . . . . . . . . . . . . . .
63 63 64 64 65 67 67 69 70 70 70 72 74 76 76 76 77 78 79 79 79 80 81 84 85 87 89 90 91 92 93 95 95 96
The Splitting Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 6.2 Splitting into Two Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 6.2.1 A General Technique of Splitting into Two Vectors . . . . 99 6.2.2 Splitting Based on the Choice of π a (t) . . . . . . . . . . . . . . . 101 6.2.3 Methods Based on the Choice of a Direction . . . . . . . . . . 102
X
Contents
6.2.4 Minimum Support Design . . . . . . . . . . . . . . . . . . . . . . . . . . 102 6.2.5 Splitting into Simple Random Sampling . . . . . . . . . . . . . . 104 6.2.6 The Pivotal Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 6.2.7 Random Direction Method . . . . . . . . . . . . . . . . . . . . . . . . . 108 6.2.8 Generalized Sunter Method . . . . . . . . . . . . . . . . . . . . . . . . . 108 6.3 Splitting into M Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 6.3.1 A General Method of Splitting into M Vectors . . . . . . . . 111 6.3.2 Brewer’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 6.3.3 Eliminatory Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 6.3.4 Till´e’s Elimination Procedure . . . . . . . . . . . . . . . . . . . . . . . 115 6.3.5 Generalized Midzuno Method . . . . . . . . . . . . . . . . . . . . . . . 117 6.3.6 Chao’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 6.4 Splitting Procedures in Brewer and Hanif . . . . . . . . . . . . . . . . . . . 120 7
More on Unequal Probability Designs . . . . . . . . . . . . . . . . . . . . . 123 7.1 Ordered Systematic Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 7.2 Random Systematic Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 7.3 Deville’s Systematic Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 7.4 Sampford Rejective Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 7.4.1 The Sampford Sampling Design . . . . . . . . . . . . . . . . . . . . . 130 7.4.2 Fundamental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 7.4.3 Technicalities for the Computation of the Joint Inclusion Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 7.4.4 Implementation of the Sampford Design . . . . . . . . . . . . . . 135 7.5 Variance Approximation and Estimation . . . . . . . . . . . . . . . . . . . 137 7.5.1 Approximation of the Variance . . . . . . . . . . . . . . . . . . . . . . 137 7.5.2 Estimators Based on Approximations . . . . . . . . . . . . . . . . 140 7.6 Comparisons of Methods with Unequal Probabilities . . . . . . . . . 142 7.7 Choice of a Method of Sampling with Unequal Probabilities . . . 145
8
Balanced Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 8.2 The Problem of Balanced Sampling . . . . . . . . . . . . . . . . . . . . . . . . 147 8.2.1 Definition of Balanced Sampling . . . . . . . . . . . . . . . . . . . . . 147 8.2.2 Particular Cases of Balanced Sampling . . . . . . . . . . . . . . . 148 8.2.3 The Rounding Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 8.3 Procedures for Equal Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . 150 8.3.1 Neyman’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 8.3.2 Yates’s Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 8.3.3 Deville, Grosbras, and Roth Procedure . . . . . . . . . . . . . . . 151 8.4 Cube Representation of Balanced Samples . . . . . . . . . . . . . . . . . . 151 8.4.1 Constraint Subspace and Cube of Samples . . . . . . . . . . . . 151 8.4.2 Geometrical Representation of the Rounding Problem . . 153 8.5 Enumerative Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 8.6 The Cube Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
Contents
8.7
8.8
8.9
8.10 9
XI
8.6.1 The Flight Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 8.6.2 Fast Implementation of the Flight Phase . . . . . . . . . . . . . 161 8.6.3 The Landing Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163 8.6.4 Quality of Balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164 Application of the Cube Method to Particular Cases . . . . . . . . . 166 8.7.1 Simple Random Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . 166 8.7.2 Stratification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 8.7.3 Unequal Probability Sampling with Fixed Sample Size . 168 Variance Approximations in Balanced Sampling . . . . . . . . . . . . . 169 8.8.1 Construction of an Approximation . . . . . . . . . . . . . . . . . . . 169 8.8.2 Application of the Variance Approximation to Stratification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 Variance Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 8.9.1 Construction of an Estimator of Variance . . . . . . . . . . . . . 173 8.9.2 Application to Stratification of the Estimators of Variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 Recent Developments in Balanced Sampling . . . . . . . . . . . . . . . . 176
An Example of the Cube Method . . . . . . . . . . . . . . . . . . . . . . . . . . 177 9.1 Implementation of the Cube Method . . . . . . . . . . . . . . . . . . . . . . . 177 9.2 The Data of Municipalities in the Canton of Ticino (Switzerland) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178 9.3 The Sample of Municipalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178 9.4 Quality of Balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Appendix: Population of Municipalities in the Canton of Ticino (Switzerland) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181 List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 List of Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 List of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194 Table of Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
1 Introduction and Overview
1.1 Purpose The aim of sampling methods is to provide information on a population by studying only a subset of it, called a sample. Sampling is the process of selecting units (households, businesses, people) from a population so that the sample allows estimating unknown quantities of the population. This book is an attempt to present a unified theory of sampling. The objective is also to describe precisely and rigorously the 46 procedures presented in this book. The R “sampling” package available on the R language Web site contains an implementation of most of the described algorithms.
1.2 Representativeness One often says that a sample is representative if it is a reduced model of the population. Representativeness is then adduced as an argument of validity: a good sample must resemble the population of interest in such a way that some categories appear with the same proportions in the sample as in the population. This theory, currently spread by the media, is, however, erroneous. It is often more desirable to overrepresent some categories of the population or even to select units with unequal probabilities. The sample must not be a reduced model of the population; it is only a tool used to provide estimates. Suppose that the aim is to estimate the production of iron in a country, and that we know that the iron is produced, on the one hand, by two huge steel companies with several thousands of workers and, on the other hand, by several hundreds of small craft-industries of less than 50 workers. Does the better design consist in selecting each unit with the same probability? Obviously, no. First, one will inquire about the production of the two biggest companies. Next, one will select a sample of small companies according to an appropriate sampling design. This simple example runs counter to the
2
1 Introduction and Overview
idea of representativeness and shows the interest of sampling with unequal probabilities. Sampling with unequal probabilities is also very commonly used in the selfweighted two-stage sampling designs. The primary units are first selected with inclusion probabilities proportional to their sizes and next, the same number of secondary units is selected in each primary unit. This design has several advantages: the sample size is fixed and the work burden can easily be balanced between the interviewers. Another definition of representativeness is that the sample must be random and all the statistical units must have a strictly positive probability of being selected. Otherwise, the sample has a problem of coverage and an unbiased estimation of some totals is impossible. This definition is obviously wise, but this term has been so overused that I think it should not be used anymore. For H´ ajek (1981), a strategy is a pair consisting of a sampling design and an estimator. A strategy is said to be representative if it allows estimating a total of the population exactly, that is, without bias and with a null variance. If the simple Horvitz-Thompson estimator is used, a strategy can be representative only if the sample automatically reproduces some totals of the population; such samples are called balanced. A large part of this book is dedicated to balanced sampling.
1.3 The Origin of Sampling Theory 1.3.1 Sampling with Unequal Probabilities Sampling methods were already mentioned in Tschuprow (1923) and Neyman (1934). Nevertheless, the first papers dedicated to sampling procedures were devoted to unequal probability sampling. Hansen and Hurwitz (1943) proposed the multinomial design with replacement with an unbiased estimator. Rapidly, it appears that unequal probability sampling without replacement is much more complicated. Numerous papers have been published, but most of the proposed methods are limited to a sample size equal to 2. Brewer and Hanif (1983) constructed a very important synthesis in which 50 methods were presented according to their date of publication, but only 20 of them really work and many of the exact procedures are very slow to apply. The well-known systematic sampling design (see Madow, 1949; Sampford, 1967) procedure remain very good solutions for sampling with unequal probabilities. 1.3.2 Conditional Poisson Sampling A very important method is Conditional Poisson Sampling (CPS), also called sampling with maximum entropy. This sampling design is obtained by selecting Poisson samples until a given sample size is obtained. Moreover, conditional Poisson sampling can also be obtained by maximizing the entropy
1.3 The Origin of Sampling Theory
3
subject to given inclusion probabilities. This sampling design appears to be natural as it maximizes the randomness of the selection of the sample. Conditional Poisson sampling was really a Holy Grail for H´ ajek (1981), who dedicated a large part of his sampling research to this question. The main problem is that the inclusion probabilities of conditional Poisson sampling change according to the sample size. H´ ajek (1981) proposed several approximations of these inclusion probabilities in order to implement this sampling design. Chen et al. (1994) linked conditional Poisson sampling with the theory of exponential families, which has paved the way for several quick implementations of the method. 1.3.3 The Splitting Technique Since the book of Brewer and Hanif (1983), several new methods of sampling with unequal probabilities have been published. In Deville and Till´e (1998), eight new methods were proposed, but the splitting method proposed in the same paper was also a way to present, sometimes more simply, almost all the existing methods. The splitting technique is thus a means to integrate the presentation of well-known methods and to make them comparable. 1.3.4 Balanced Sampling The interest of balanced sampling was already pointed out more than 50 years ago by Yates (1946) and Thionet (1953). Several partial solutions of balanced sampling methods have been proposed by Deville et al. (1988), Ardilly (1991), Deville (1992), Hedayat and Majumdar (1995), and Valliant et al. (2000). Royall and Herson (1973a,b) and Scott et al. (1978) discussed the importance of balanced sampling in order to protect the inference against a misspecified model. They proposed optimal estimators under a regression model. In this model-based approach, the optimality is conceived only with respect to the regression model without taking into account the sampling design. Nevertheless, these authors come to the conclusion that the sample must be balanced but not necessarily random. Curiously, they developed this theory while there still did not exist any general method for selecting balanced samples. Recently, Deville and Till´e (2004) proposed a general procedure, the cube method, that allows the selection of random balanced samples on several balancing variables, with equal or unequal inclusion probabilities in the sense that the Horvitz-Thompson estimator is equal or almost equal to the population total of the balancing variables. Deville and Till´e (2005) have also proposed an approximation of the variance for the Horvitz-Thompson estimator in balanced sampling with large entropy.
4
1 Introduction and Overview
1.4 Scope of Application Sampling methods can be useful when it is impossible to examine all the units of a finite population. When a sampling frame or register is available, it is possible to obtain balancing information on the units of interest. In most cases, it is possible to take advantage of this information in order to increase the accuracy. In business statistics, registers are generally available. Several countries have a register of the population. Nevertheless, in the two-stage sampling design, the primary units are usually geographic areas, for which a large number of auxiliary variables are generally available. The Institut National de la Statistique et de l’Analyse de l’Information (INSEE, France) has selected balanced samples for the most important statistical projects. In the redesigned census in France, a fifth of the municipalities with fewer than 10,000 inhabitants are sampled each year, so that after five years all the municipalities will be selected. All the households in these municipalities are surveyed. The five samples of municipalities are selected with equal probabilities using balanced sampling on a set of demographic variables. This methodology ensures the accuracy of the estimators based upon the five groups of rotation. The selection of the primary units (areas) for INSEE’s new master sample for household surveys is done with probabilities proportional to the number of dwellings using balanced sampling. Again, the balancing variables are a set of demographic and economic variables. The master sample is used over ten years to select all the households samples. This methodology ensures a better accuracy of the estimates depending on the master sample.
1.5 Aim of This Book In view of the large number of publications, a synthesis is probably necessary. Recently, solutions have been given for two important problems in survey sampling: the implementation of conditional Poisson sampling and a rapid method for selecting balanced samples, but the research is never closed. A very open field is the problem of the co-ordination of samples in repeated surveys. The challenge of this book is to propose at the same time a unified theory of the sampling methods and tools that are directly applicable to practical situations. The samples are formally defined as random vectors whose components denote the number of times a unit is selected in the sample. From the beginning, the notions of sampling design and sampling algorithm are differentiated. A sampling design is nothing more than a multivariate discrete distribution, whereas a sampling algorithm is a way of implementing a sampling design. Particular stress is given to a geometric representation of the sampling methods.
1.6 Outline of This Book
5
This book contains a very precise description of the sampling procedures in order to be directly applicable in real applications. Moreover, an implementation in the R language available on the website of the Comprehensive R Archive Network allows the reader to directly use the described methods (see Till´e and Matei, 2005).
1.6 Outline of This Book In Chapter 2, a particular notation is defined. The sample is represented by a vector of RN whose components are the number of times each unit is selected in the sample. The interest of this notation lies in the possibility of dealing with sampling in the same way with or without replacement. A sample therefore becomes a random vector with positive integer values and the sampling designs are discrete multivariate distributions. In Chapter 3, general algorithms are presented. In fact, any particular sampling design can be implemented by means of several algorithms. Chapter 4 is devoted to simple random sampling. Again, an original definition is proposed. From a general expression, all the simple random sampling designs can be derived by means of change of support, a support being a set of samples. Chapter 5 is dedicated to exponential sampling designs. Chen et al. (1994) have linked these sampling designs to random vectors with positive integer values and exponential distribution. Again, changes of support allow defining the most classic sampling designs, such as Poisson sampling, conditional Poisson sampling, and multinomial sampling. The exact implementation of the conditional Poisson design, that could also be called “exponential design with fixed sample size”, was, since the book of H´ ajek, a Holy Grail that seemed inaccessible for reasons of combinatory explosion. Since the publication of the paper of Chen et al. (1994), important progress has been achieved in such a way that there now exist very fast algorithms that implement this design. Chapter 6 is devoted to the splitting method proposed by Deville and Till´e (1998). This method is a process that allows constructing sampling procedures. Most of the sampling methods are best presented in the form of the splitting technique. A sequence of methods is presented, but it is possible to construct many other ones with the splitting method. Some well-known procedures, such as the Chao (1982) or the Brewer (1975) methods, are presented only by means of the splitting scheme because they are more comparable with this presentation. In Chapter 7, are presented some sampling methods with unequal probabilities that are not exponential and that cannot be presented by means of the splitting scheme: systematic (see Madow, 1949) sampling, the Deville (1998) systematic method, and the Sampford (1967) method. The choice of the method is next discussed. When the same inclusion probabilities are used, it is possible to proof that none of the methods without replacement provides a better accuracy than the other ones. The choice of the method cannot thus
6
1 Introduction and Overview
be done in function of the accuracy, but essentially in function of practical considerations. The problem of variance estimation is also discussed. Two approximations are constructed by means of a simple sum of squares. Several sampling methods are compared to these approximations. One can notice that the variance of the Sampford method and the exponential procedure are very well adjusted by one of these approximations, which provides a guideline for the choice of a sampling method and of a variance estimator. In Chapter 8, methods of balanced sampling are developed. After a presentation of the methods for equal inclusion probabilities, the cube method is described. This method allows selecting balanced samples on several dozens of variables. Two algorithms are presented, and the question of variance estimation in balanced sampling is developed. Finally, in Chapter 9, the cube method is applied in order to select a sample of municipalities in the Swiss canton of Ticino.
2 Population, Sample, and Sampling Design
2.1 Introduction In this chapter, some basic concepts of survey sampling are introduced. An original notation is used. Indeed, a sample without replacement is usually defined as a subset of a finite population. However, in this book, we define a sample as a vector of indicator variables. Each component of this vector is the number of times that a unit is selected in the sample. This formalization is, however, not new. Among others, it has been used by Chen et al. (1994) for exponential designs and by Deville and Till´e (2004) for developing the cube method. This notation tends to be obvious because it allows defining a sampling design as a discrete multivariate distribution (Traat, 1997; Traat et al., 2004) and allows dealing with sampling with and without replacement in the same way. Of particular importance is the geometrical representation of a sampling design that is used in Chapter 8, which is dedicated to balanced sampling.
2.2 Population and Variable of Interest A finite population is a set of N units {u1 , . . . , uk , . . . , uN }. Each unit can be identified without ambiguity by a label. Let U = {1, . . . , k, . . . , N } be the set of these labels. The size of the population N is not necessarily known. The aim is to study a variable of interest y that takes the value yk for unit k. Note that the yk ’s are not random. The objective is more specifically to estimate a function of interest T of the yk ’s: T = f (y1 · · · yk · · · yN ).
8
2 Population, Sample, and Sampling Design
The most common functions of interest are the total Y = yk , k∈U
the population size N=
1,
k∈U
the mean Y =
1 yk , N k∈U
the variance σy2 =
2 1 1 2 (yk − y ) yk − Y = N 2N 2 k∈U
k∈U ∈U
and the corrected variance 2 1 1 2 yk − Y = (yk − y ) . Vy2 = N −1 2N (N − 1) k∈U
k∈U ∈U
2.3 Sample 2.3.1 Sample Without Replacement A sample without replacement is denoted by a column vector s = (s1 · · · sk · · · sN ) ∈ {0, 1}N ,
where sk =
1 if unit k is in the sample 0 if unit k is not in the sample,
for all k ∈ U. The sample size is n(s) =
sk .
k∈U
Note that the null vector s = (0 · · · 0 · · · 0) is a sample, called the empty sample and that the vector of ones s = (1 · · · 1 · · · 1) is also a sample, called the census.
2.4 Support
9
2.3.2 Sample With Replacement The same kind of notation can be used for samples with replacement, which can also be denoted by a column vector s = (s1 · · · sk · · · sN ) ∈ NN , where N = {0, 1, 2, 3, . . . } is the set of natural numbers and sk is the number of times that unit k is in the sample. Again, the sample size is n(s) = sk , k∈U
and, in sampling with replacement, we can have n(s) > N.
2.4 Support Definition 1. A support Q is a set of samples. Definition 2. A support Q is said to be symmetric if, for any s ∈ Q, all the permutations of the coordinates of s are also in Q. Some particular symmetric supports are used. Definition 3. The symmetric support without replacement is defined by S = {0, 1}N . Note that card(S) = 2N . The support S can be viewed as the set of all the vertices of a hypercube of RN (or N -cube), where R is the set of real numbers, as shown for a population size N = 3 in Figure 2.1.
(011)
(111)
(010)
(110)
(001) (101)
(000)
(100)
Fig. 2.1. Possible samples without replacement in a population of size N = 3
10
2 Population, Sample, and Sampling Design
Definition 4. The symmetric support without replacement with fixed sample size is Sn = s ∈ S sk = n . k∈U Note that card(Sn ) = N n . Figure 2.2 shows an example of S2 in a population of size N = 3. (011)
(111)
(010)
(110)
(001) (101)
(000)
(100)
Fig. 2.2. The support S2 in a population of size N = 3
Definition 5. The symmetric support with replacement is R = NN , where N is the set of natural numbers. The full support with replacement R is countably infinite. As shown in Figure 2.3, support R is the lattice of all the vector of NN . Definition 6. The symmetric support with replacement of fixed size n is defined by sk = n . Rn = s ∈ R k∈U
Result 1. The size of Rn is card(Rn ) =
N +n−1 n
.
(2.1)
Proof. (by induction) If G(N, n) denotes the number of elements in Rn in a population of size N, then G(1, n) = 1. Moreover, we have the recurrence relation n G(N + 1, n) = G(N, i). (2.2) i=0
Expression (2.1) satisfies the recurrence relation (2.2).
2
2.4 Support
(021) (020)
(120)
(011) (010)
(110)
(001) (000)
(100)
(121)
11
(221)
(220)
(111)
(211)
(210)
(101)
(201)
(200)
Fig. 2.3. Geometrical representation of support R
Figure 2.4 shows the set of samples with replacement of size 2 in a population of size 3.
(020)
(011) (110)
(002) (101) (200)
Fig. 2.4. Samples with replacement of size 2 in a population of size 3
The following properties follow directly: 1. 2. 3. 4. 5.
S, Sn , R, Rn , are symmetric, S ⊂ R, The set {S0 , . . . , Sn , . . . , SN } is a partition of S, The set {R0 , . . . , Rn , . . . , RN , . . . } is an infinite partition of R, Sn ⊂ Rn , for all n = 0, . . . , N.
12
2 Population, Sample, and Sampling Design
2.5 Convex Hull, Interior, and Subspaces Spanned by a Support A support can be viewed as a set of points in RN . In what follows, several subsets of RN are defined by means of the support: the convex hull, its interior, and several subspaces generated by the support. This geometrical vision is especially useful in Chapter 5 dedicated to exponential design and in Chapter 8 dedicated to balanced sampling. We refer to the following definitions. Definition 7. Let s1 , . . . , si , . . . , sI be the enumeration of all the samples of Q. The convex hull of the support Q is the set I I λi si , λ1 , . . . , λI , of R+ , such that λi = 1 , Conv Q = i=1
i=1
where R+ = [0, ∞) is the set of nonnegative real numbers. Definition 8. Let s1 , . . . , si , . . . , sI be the enumeration of all the samples of Q. The affine subspace spanned by a support is the set I I λi si , λ1 , . . . , λI , of R, such that λi = 1 . Aff Q = i=1
i=1
→ − Definition 9. The direction Q of the affine subspace spanned by a support Q (the direction of the support for short) is the linear subspace spanned by all the differences si − sj , for all si , sj ∈ Q. → − Note that Q is a linear subspace. Definition 10. Let s1 , . . . , si , . . . , sI be the enumeration of all the samples of Q. The interior of Conv Q in the affine subspace spanned by Q is defined by I I ◦ λi si λi > 0, for all i = 1, . . . , I, and λi = 1 . Q= i=1
i=1
◦
In short, afterwards Q is called the interior of Q. Definition 11. The invariant subspace spanned by the support is defined by
Invariant Q = u ∈ RN |u (s1 − s2 ) = 0, for all s1 , s2 ∈ Q . ◦
Remark 1. Q⊂ Conv Q ⊂ Aff Q. → − Remark 2. Invariant Q and Q are two linear subspaces. Invariant Q is the → − orthogonal subspace of Q.
2.5 Convex Hull, Interior, and Subspaces Spanned by a Support
13
Table 2.1 gives the list of the most common supports and their invariant subspace, direction, affine subspace, convex hull, and interior, where R∗+ = {x ∈ R|x > 0} is the set of positive real numbers. Table 2.1. Invariant, direction, affine subspace, convex hull, and interior of the supports R, Rn , S, and Sn R Invariant
{0}
Direction
RN
Aff
RN
Conv
(R+ )N
Interior
N (R∗ +)
Rn S Sn x ∈ RN |x = a1, for all a ∈ R {0} x ∈ RN |x = a1, for all a ∈ R N N u ∈ RN u ∈ RN ui = 0 RN ui = 0 k=1 k=1 N N v ∈ RN v ∈ RN vi = n RN vi = n i=1 i=1 N N v ∈ [0, n]N v ∈ [0, 1]N vi = n [0, 1]N vi = n i=1 i=1 N N N N N v ∈]0, n[ v ∈]0, 1[ vi = n ]0, 1[ vi = n i=1
i=1
Example 1. Let U = {1, 2, 3, 4} and ⎧⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎫ 1 1 0 0 ⎪ ⎪ ⎪ ⎨⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎪ ⎬ 0 0 1 ⎟ , ⎜ ⎟ , ⎜ ⎟ , ⎜ 1⎟ . Q= ⎜ ⎝1⎠ ⎝0⎠ ⎝1⎠ ⎝0⎠⎪ ⎪ ⎪ ⎪ ⎩ ⎭ 0 1 0 1 Then
Aff Q = u ∈ RN |u1 + u2 = 1 and u3 + u4 = 1 ,
Conv Q = u ∈ [0, 1]N |u1 + u2 = 1 and u3 + u4 = 1 , ◦
Q= u ∈]0, 1[N |u1 + u2 = 1 and u3 + u4 = 1 , ⎧ ⎛ ⎞ ⎫ ⎛ ⎞ 1 0 ⎪ ⎪ ⎪ ⎪ ⎨ ⎜ ⎟ ⎬ ⎜0⎟ 1 ⎟ + b ⎜ ⎟ , for all a, b ∈ R , Invariant Q = a ⎜ ⎝ ⎠ ⎝ ⎠ 0 1 ⎪ ⎪ ⎪ ⎪ ⎩ ⎭ 0 1
and
⎧ ⎛ ⎞ ⎫ ⎛ ⎞ 1 0 ⎪ ⎪ ⎪ ⎪ ⎬ ⎜0⎟ → ⎨ ⎜ − −1⎟ ⎜ ⎟ ⎜ ⎟ Q = a ⎝ ⎠ + b ⎝ ⎠ , for all a, b ∈ R . 0 1 ⎪ ⎪ ⎪ ⎪ ⎩ ⎭ 0 −1
14
2 Population, Sample, and Sampling Design
2.6 Sampling Design and Random Sample Definition 12. A sampling design p(.) on a support Q is a multivariate probability distribution on Q; that is, p(.) is a function from support Q to ]0, 1] such that p(s) > 0 for all s ∈ Q and p(s) = 1. s∈Q
Definition 13. A sampling design with support Q is said to be without replacement if Q ⊂ S. Definition 14. A sampling design with support Q is said to be of fixed sample size n if Q ⊂ Rn . Remark 3. Because S can be viewed as the set of all the vertices of a hypercube, a sampling design without replacement is a probability measure on all these vertices. Definition 15. A sampling design of support Q is said to be with replacement if Q\S is nonempty. Definition 16. If p(.) is a sampling design without replacement, the complementary design pc (.) of a sampling design p(c) of support Q is pc (1 − s) = p(s), for all s ∈ Q, where 1 is the vector of ones of RN . Definition 17. A random sample S ∈ RN with the sampling design p(.) is a random vector such that Pr(S = s) = p(s), for all s ∈ Q, where Q is the support of p(.).
2.7 Reduction of a Sampling Design With Replacement Let S = (S1 · · · Sk · · · SN ) be a sample with replacement. The reduction function r(.) suppresses the information about the multiplicity of the units, as follows. If S∗ = r(S),
then Sk∗
=
1 if Sk > 0 0 if Sk = 0.
(2.3)
2.8 Expectation and Variance of a Random Sample
15
The reduction function allows constructing a design without replacement from a design with replacement. If p(.) is a sampling design with support R, then it is possible to derive a design p∗ (.) on S in the following way: p(s), p∗ (s∗ ) = s∈R|s∗ =r(s)
for all s ∈ S. Example 2. Let p(.) be a sampling design with replacement of fixed size n = 2 on U = {1, 2, 3} such that p((2, 0, 0) ) =
1 , 9 2 p((1, 1, 0) ) = , 9 then p∗ (.) is given by
p((0, 2, 0) ) =
1 , 9 2 p((1, 0, 1) ) = , 9
p((0, 0, 2) ) =
1 , 9 2 p((1, 1, 0) ) = , 9
p((0, 1, 0) ) =
1 , 9 2 p((1, 0, 1) ) = , 9
p((0, 0, 1) ) =
p((1, 0, 0) ) =
1 , 9 2 p((0, 1, 1) ) = ; 9 1 , 9 2 p((0, 1, 1) ) = . 9
2.8 Expectation and Variance of a Random Sample Definition 18. The expectation of a random sample S is µ = E(S) = p(s)s. s∈Q
Remark 4. Because µ is a linear convex combination with positive coefficient ◦
of the elements of Q, µ ∈Q . Remark 5. Let C be the N × I matrix of all the centered samples of Q; that is C = (s1 − µ · · · sI − µ), (2.4) where I = cardQ. Then Ker C = Invariant Q, where Ker C = {u ∈ RN |C u = 0}. The joint expectation is defined by p(s)sk s . µk = s∈Q
Finally, the variance-covariance operator is p(s)(s − µ)(s − µ) = [µk − µk µ ]. Σ = [Σk ] = var(S) = s∈Q
16
2 Population, Sample, and Sampling Design
Result 2.
µk = E[n(S)],
k∈U
and, if the sampling design is of fixed sample size n(S), µk = n(S). k∈U
Proof. Because
Sk = n(S),
k∈U
we have
E(Sk ) =
k∈U
µk = E [n(S)] .
2
k∈U
Result 3. Let Σ = [Σk ] be the variance-covariance operator. Then Σk = E [n(S)(S − µ )] , for all ∈ U, k∈U
and, if the sampling design is of fixed sample size n, Σk = 0, for all ∈ U.
(2.5)
k∈U
Proof. Because Σk = (µk − µk µ ) = [E(Sk S ) − E(Sk )E(S )] k∈U
k∈U
k∈U
= E [n(S)(S − µ )] . If var[n(S)] = 0, then Σk = E [n(S)(S − µ )] = n(S)E (S − µ ) = 0.
2
k∈U
Result 4. Let p(.) be a sampling design with support Q. Then Ker Σ = Invariant Q. Proof. Let C denote the matrix of the centered samples defined in (2.4). Because Σ = CPC where P = diag[p(s1 ) · · · p(sI )], we have that, if u ∈ Invariant Q, C u = 0, thus Σu = 0; that is, Ker Σ ⊂ Invariant Q. Moreover, because P is of full rank, C and Σ have the same rank, which implies that Ker Σ and Ker C have the same dimension. Thus, Ker Σ = Invariant Q. 2
2.9 Inclusion Probabilities
17
Remark 6. Let the image of Σ be
Im(Σ) = x ∈ RN | there exists u ∈ RN such that Σu = x . → − Because Im(Σ) is the orthogonal of KerΣ, Im(Σ) = Q. Example 3. If the sample is of fixed sample size, 1 ∈ Invariant Q, Σ1 = 0 where 1 = (1 · · · 1 · · · 1) . Result 4 is thus a generalization of Expression (2.5).
2.9 Inclusion Probabilities Definition 19. The first-order inclusion probability is the probability that unit k is in the random sample πk = Pr(Sk > 0) = E[r(Sk )], where r(.) is the reduction function. Moreover, π = (π1 · · · πk · · · πN ) denotes the vector of inclusion probabilities. If the sampling design is without replacement, then π = µ. Definition 20. The joint inclusion probability is the probability that unit k and are together in the random sample πk = Pr(Sk > 0 and S > 0) = E [r(Sk )r(S )] , with πkk = πk , k ∈ U. Let Π = [πk ] be the matrix of joint inclusion probabilities. Moreover, we define ∆ = Π − ππ . Note that ∆kk = πk (1 − πk ), k ∈ U. If the sample is without replacement, then ∆ = Σ.
Result 5.
πk = E {n[r(S)]} ,
k∈U
and
∆k = E {n[r(S)] (r(S ) − π )} , for all ∈ U.
k∈U
Moreover, if var {n[r(S)]} = 0 then ∆k = 0, for all ∈ U. k∈U
18
2 Population, Sample, and Sampling Design
The proof is the same as for Results 2 and 3. When the sample is without replacement and of fixed sample size, then S = r(S), and Result 5 becomes πk = n(S), k∈U
and
∆k = 0, for all ∈ U.
k∈U
2.10 Computation of the Inclusion Probabilities Suppose that the values xk of an auxiliary variable x are known for all the units of U and that the xk > 0, for all k ∈ U. When the xk ’s are approximately proportional to the values yk ’s of the variable of interest, it is then interesting to select the units with unequal probabilities proportional to the xk in order to get an estimator of Y with a small variance. Unequal probability sampling is also used in multistage selfweighted sampling designs (see, for instance, S¨ arndal et al., 1992, p. 141). To implement such a sampling design, the inclusion probabilities πk are computed as follows. First, compute the quantities nx k ∈U
x
,
(2.6)
k = 1, . . . , N. For units for which these quantities are larger than 1, set πk = 1. Next, the quantities are recalculated using (2.6) restricted to the remaining units. This procedure is repeated until each πk is in ]0, 1]. Some πk are 1 and the others are proportional to xk . This procedure is formalized in Algorithm 2.1. More formally, define x k (2.7) h(z) = min z , 1 , X where X =
k∈U
k∈U
xk . Then, the inclusion probabilities are given by xk , k ∈ U. πk = min 1, h−1 (n) X
(2.8)
The selection of the units with πk = 1 is trivial, therefore we can consider that the problem consists of selecting n units without replacement from a population of size N with fixed πk , where 0 < πk < 1, k ∈ U, and πk = n. k∈U
2.11 Characteristic Function of a Sampling Design
19
Algorithm 2.1 Computation of the inclusion probabilities Definition f : vectors of N Integers; π: vectors of N Reals; X Real, m, n, m1 Integer; For k = 1, . . . , N do fk = 0 ; EndFor; m = 0; m1 = −1; While (m1 = m), do m1 = m; X= N k=1 xk (1 − fk ); c = n − m; For k = 1, . . . , N do If fk = 0 then πk = cxk /X; If πk ≥ 1 then fk = 1; EndIf; Else πk = 1; EndIf; EndFor; m= N k=1 fk ; EndWhile.
2.11 Characteristic Function of a Sampling Design A sampling design can thus be viewed as a multivariate distribution, which allows defining the characteristic function of a random sample. Definition 21. The characteristic function φ(t) from RN to C of a random sample S with sampling design p(.) on Q is defined by φS (t) = eit s p(s), t ∈ RN , (2.9) s∈Q
where i =
√
−1, and C is the set of the complex numbers.
In some cases (see Chapters 4 and 5), the expression of the characteristic function can be significantly simplified in such a way that the sum over all the samples of the support does not appear anymore. When the characteristic function can be simplified, the sampling design is more tractable. If φ (t) denotes the vector of first derivatives and φ (t) the matrix of second derivatives of φ(t), then we have φ (0) = i sp(s) = iµ, s∈Q
and
φ (0) = −
s∈Q
ss p(s) = −(Σ + µµ ).
20
2 Population, Sample, and Sampling Design
2.12 Conditioning a Sampling Design A sampling design can be conditioned with respect to a support. Definition 22. Let Q1 and Q2 be two supports such that Q2 ⊂ Q1 and p1 (.) a sampling design on Q1 . Then, the conditional design of Q1 with respect to Q2 is given by p1 (s) , for all s ∈ Q2 . s∈Q2 p1 (s)
p1 (s|Q2 ) =
A sampling design p(.) can also be conditioned with respect to a particular value ak for Sk p(s) , p(s|Sk = ak ) = s|sk =ak p(s) for all s such that Sk = ak . Finally, a sampling design can also be conditioned with respect to particular values ak for a subset A ⊂ U of units in the sample p(s) . (Sk = ak ) = p s s| (sk =ak ) p(s) k∈A
k∈A
for all s such that sk = ak , k ∈ A.
2.13 Observed Data and Consistency Definition 23. The set D = {(Vk , Sk ), k ∈ U } is called the observed data, where Vk = yk Sk . Also, let d denote a possible value for D d = {(vk , sk ), k ∈ U }, where vk = yk sk . The set of possible values for the observed data is X = {d|s ∈ S, y ∈ RN }. The following definition is necessary to define the probability distribution of D. Definition 24. A possible value d is said to be consistent with a particular ∗ ∗ vector yN = (y1∗ · · · yk∗ · · · yN ) if and only if yk = yk∗ , for all k such that sk > 0.
2.14 Statistic and Estimation
21
Because Pr(D = d; y) = Pr(D = d and S = s; y) = Pr(D = d|S = s; y)Pr(S = s), 1 if d is consistent with y Pr(D = d|S = s; y) = 0 otherwise. The probability distribution of D can be written p(s) if d is consistent with y pY (d) = Pr(D = d; y) = 0 otherwise.
(2.10)
2.14 Statistic and Estimation Definition 25. A statistic G is a function of the observed data: G = u(D). The expectation of G is E(G) =
p(s)G(s),
s∈Q
where G(s) is the value taken by statistic G on sample s, and Q is the support of the sampling design. The variance of a statistic G is defined by 2
var(G) = E [G − E(G)] . Definition 26. An estimator T! of a function of interest T is a statistic used to estimate it. The bias of T! is defined by B(T!) = E(T!) − T, and the mean square error by 2 MSE(T!) = E T! − T = var(T!) + B2 (T!). Definition 27. An estimator Y!L of Y is said to be linear if it can be written wk yk Sk , Y!L1 = w0 + k∈U
or
Y!L2 = w0 +
wk yk r(Sk ),
k∈U
where wk can eventually depend on S and thus can be random.
22
2 Population, Sample, and Sampling Design
Result 6. Let p(.) be a sampling design such that µk > 0, for all k ∈ U. A linear estimator Y!L1 of Y is unbiased if and only if E(w0 ) = 0
and
E(wk Sk ) = 1, for all k ∈ U.
A linear estimator Y!L2 of Y is unbiased if and only if E(w0 ) = 0
and
E[wk r(Sk )] = 1, for all k ∈ U.
Because the yk ’s are not random, the proof is obvious.
2.15 Sufficient Statistic In order to deal with the estimation problem, one could use the technique of sufficiency reduction. We refer to the following definition. Definition 28. A statistic G = u(D) is said to be sufficient for vector y = (y1 · · · yN ) if and only if the conditional distribution of D given that G = g does not depend on y = (y1 · · · yN ) (as long as this conditional probability exists). If Pr(D = d|G = g; y) is the conditional probability, a statistic G is sufficient if Pr(D = d|G = g; y) is constant with respect to y as long as Pr(G = g; y) > 0. A sample with replacement can always be reduced to a sample without replacement of S by means of the reduction function r(.) given in Expression (2.3) that suppresses the information about the multiplicity of the units. A sampling design p(.) with support Q ⊂ R can also be reduced to a sampling design p∗ (.) without replacement: p∗ (s∗ ) = p(s). s∈Q|r(s)=s∗
Similarly, if the sample is with replacement, the observed data can also be reduced using function u∗ (.), where
" Vk u∗ (D) = , r(Sk ) , k ∈ U = {(yk r(Sk ), r(Sk )) , k ∈ U } 1 − r(Sk ) + Sk and Vk = yk Sk . Theorem 1 (see Basu and Ghosh, 1967; Basu, 1969; Cassel et al., 1993, p. 36; Thompson and Seber, 1996, p. 35). For all sampling designs with support Q ⊂ R, D∗ = u∗ (D) is a sufficient statistic for y. Proof. Consider the conditional probability: Pr(D = d|D∗ = d∗ ; y) =
Pr(D = d and D∗ = d∗ ; y) . Pr(D∗ = d∗ ; y)
2.15 Sufficient Statistic
23
This probability is defined only if Pr(D∗ = d∗ ; y) > 0, which implies that D∗ is consistent with y. Two cases must be distinguished. 1. If u∗ (d) = d∗ , then by (2.10) we obtain Pr(D = d and D∗ = d∗ ; y) = Pr(D = d; y) = pY (d) = p(s). 2. If u∗ (d) = d∗ , then Pr(D = d and D∗ = d∗ ; y) = 0. If d∗ is consistent with y, we have Pr(D = d; y) = p(s), and thus ∗
∗
Pr(D = d|D = d ; y) =
p(s)/p∗ (s∗ ) 0
where r(s) = s∗ and
p∗ (s∗ ) =
if d∗ = u∗ (d) u∗ (d), if d∗ =
p(s).
s∈Q|r(s)=s∗
This conditional distribution does not depend on y and u∗ (D) is thus sufficient. 2 The factorization theorem allows for the identification of a sufficient statistic: Theorem 2 (see Basu and Ghosh, 1967; Basu, 1969; Cassel et al., 1993, p. 36). For all sampling designs on Q ⊂ R then G = u(D) is sufficient for y if and only if pY (d) = g[u(d), y]h(d), where the function h(d) does not depend on y and g(.) only depends on u(d). Proof. Define the indicator function 1 if d is consistent with y δ(d, y) = 0 otherwise. Expression (2.10) can thus be written: pY (d) = p(s)δ(d; y).
(2.11)
If G = u(D) is a sufficient statistic and Pr[G = u(d); y] > 0, we have Pr[D = d|G = u(d); y] = h(d), where h(d) does not depend on y. We thus have pY (d) = Pr(D = d and G = u(d); y) = Pr(G = u(d); y)h(d) = g[u(d), y]h(d), where g[u(d), y] only depends on u(d).
2
24
2 Population, Sample, and Sampling Design
The reduction principle consists of retaining from the observed data a minimum of information in the following way: Definition 29. A statistic is said to be a minimum sufficient statistic if it can be written as a function of all the sufficient statistics. Theorem 3 (see Basu and Ghosh, 1967; Basu, 1969; Cassel et al., 1993, p. 37; Thompson and Seber, 1996, p. 38). For all sampling designs with replacement, the statistic D∗ = u∗ (D) (suppression of the information about the multiplicity of the units) is minimum sufficient. Proof. A statistic u(.) can define a partition of X denoted Pu . This partition is such that if u(d1 ) = u(d2 ) then d1 and d2 are in the same subset of the partition. A statistic u∗ (.) is thus minimum sufficient if, for every sufficient statistic u(.), each subset of the partition Pu is included in a subset of the partition Pu∗ . Let u(.) be any sufficient statistic such that u(d1 ) = u(d2 ), which means that d1 and d2 are in the same subset of the partition Pu . We show that it implies that they are in the same subset of the partition Pu∗ . Because u(.) is sufficient, by Theorem 2, we have pY (d1 ) = g [u(d1 ), y] h(d1 ) and pY (d2 ) = g [u(d2 ), y] h(d2 ). Because g[u(d1 ), y] = g[u(d2 ), y] and u(d1 ) = u(d2 ), we obtain pY (d2 ) pY (d1 ) = . h(d1 ) h(d2 ) Finally, by (2.11), we have p(s1 )δ(d1 ; y) p(s2 )δ(d2 ; y) = , h(d1 ) h(d2 ) which is true if and only if δ(d1 ; y) = δ(d2 ; y). In this case, r(s1 ) = r(s2 ) and u∗ (d1 ) = u∗ (d2 ). Thus, d1 and d2 are in the same subset of the partition Pu∗ . Because this result is true for any sufficient statistic u(.), the statistic u∗ (.) is minimum sufficient. 2 This result shows that the information about the multiplicity of the units can always be suppressed from the observed data. The Rao-Blackwell theorem shows that any estimator can be improved by means of a minimum sufficient statistic.
2.15 Sufficient Statistic
25
Theorem 4 (of Rao-Blackwell, see Cassel et al., 1993, p. 40). Let p(s) be a sampling design with replacement and an estimator T! (eventually biased) of ∗ ∗ !∗ ! ∗ T where = u (D). If T = E(T |D ), then D (i) E T!∗ = E T! , 2 (ii) MSE T! = MSE T!∗ + E T!∗ − T! , (iii) MSE T!∗ ≤ MSE T! . Proof. (i) is obvious because E(T!) = EE(T!|D) = E(T!∗ ). Next, for (ii), we have 2 MSE T! = E T! − T!∗ + T!∗ − T 2 2 # $ = E T! − T!∗ + E T!∗ − T − 2E T! − T!∗ T!∗ − T . The third term is null because $ # $ # = E T!∗ − T E T! − T!∗ |D T! − T!∗ = 0. E T!∗ − T 2
Finally, (iii) comes directly from (ii).
The classical example of Rao-Blackwellization is the estimation of a total in simple random sampling with replacement and is developed in Section 4.6, page 53. The result about sufficiency is interesting but weak. The information about the multiplicity of the units can always be suppressed, but the sufficiency reduction does not permit the construction of an estimator. Because it is impossible to identify an estimator by means of the technique of reduction by sufficiency, we are restricted to the use of linear unbiased estimators of the totals defined on page 21. Consider two cases. • The first estimator takes into account the multiplicity of the units: wk yk Sk , Y!L1 = w0 + k∈U
where E(w0 ) = 0 and E(wk Sk ) = 1, for all k ∈ U. If the wk ’s are not random, the only solution is: w0 = 0 and wk = 1/E(Sk ) = 1/µk , for all k ∈ U, which gives the Hansen-Hurwitz estimator. • The second estimator depends on the sample only through the reduction function: Y!L2 = w0 + wk yk r(Sk ), k∈U
where E(w0 ) = 0 and E[wk r(Sk )] = 1, for all k ∈ U. If the wk ’s are not random, the only solution is: w0 = 0 and wk = 1/E[r(Sk )] = 1/πk , for all k ∈ U, which gives the Horvitz-Thompson estimator.
26
2 Population, Sample, and Sampling Design
2.16 The Hansen-Hurwitz (HH) Estimator 2.16.1 Estimation of a Total Definition 30. The Hansen-Hurwitz estimator (see Hansen and Hurwitz, 1943) of Y is defined by Sk y k , Y!HH = µk k∈U
where µk = E(Sk ), k ∈ U. Result 7. If µk > 0, for all k ∈ U, then Y!HH is an unbiased estimator of Y . Proof. If µk > 0, for all k ∈ U, then E(Y!HH ) =
E(Sk )yk = yk = Y. µk
k∈U
2
k∈U
2.16.2 Variance of the Hansen-Hurwitz Estimator The variance of the Hansen-Hurwitz estimator is var1 (Y!HH ) =
yk y Σk . µk µ
(2.12)
k∈U ∈U
Result 8. When the design is of fixed sample size, the variance can also be written:
2 1 yk y var2 (Y!HH ) = − − Σk . (2.13) 2 µk µ k∈U ∈U =k
Proof. We have −
2
2 1 yk y 1 yk y − Σk = − − Σk 2 µk µ 2 µk µ k∈U ∈U =k
=−
k∈U ∈U
y2
k∈U ∈U
k Σk µ2k
−
yk y Σk µk µ
=
yk y y2 k Σk − Σk . µk µ µ2k
k∈U ∈U
k∈U
∈U
When the sampling design has a fixed sample size (see Result 3, page 16), Σk = 0, ∈U
which proves Result 8.
2
2.16 The Hansen-Hurwitz (HH) Estimator
27
2.16.3 Variance Estimation From expressions (2.12) and (2.13), it is possible to construct two variance estimators. For general sampling designs, we have: var % 1 (Y!HH ) =
Sk S yk y Σk . µk µ µk
(2.14)
k∈U ∈U
For sampling designs with fixed sample size, we have: 1 var % 2 (Y!HH ) = − 2
2 yk y Sk S Σk − . µk µ µk
(2.15)
k∈U ∈U =k
Both estimators are unbiased if µk > 0 for all k, ∈ U. Indeed, y y Σ k k E var % 1 (Y!HH ) = E(Sk S ) = var1 (Y!HH ), µk µ µk k∈U ∈U
and
2 Σk 1 yk y ! E var % 2 (YHH ) = − − E(Sk S ) = var2 (Y!HH ). 2 µk µ µk
k∈U ∈U =k
By developing var % 2 (Y!HH ), we get the following result. Result 9. var % 1 (Y!HH ) = var % 2 (Y!HH ) +
Sk y 2 S Σk k . µ2k µk
k∈U
∈U
Note that, with fixed sample size, estimator (2.14) should never be used. Indeed, if yk = µk , k ∈ U and the sampling design has a fixed sample size, we have var(Y!HH ) = 0, var % 2 (Y!HH ) = 0, but var % 1 (Y!HH ) =
k∈U
which is not generally equal to 0.
Sk
S Σk , µk
∈U
28
2 Population, Sample, and Sampling Design
2.17 The Horvitz-Thompson (HT) Estimator 2.17.1 Estimation of a Total The Horvitz-Thompson estimator(see Horvitz and Thompson, 1952) is defined by r(Sk )yk ˇ, = r(S) y Y!HT = πk k∈U
ˇ = (y1 /π1 · · · yk /πk · · · yk /πk ). When sampling is without rewhere y placement, the Horvitz-Thompson estimator is equal to the Hansen-Hurwitz estimator. 2.17.2 Variance of the Horvitz-Thompson Estimator The variance of the Horvitz-Thompson estimator is yk y ∆k ˇ ∆ˇ var1 (Y!HT ) = =y y. πk π
(2.16)
k∈U ∈U
When the size of the reduced sample n[r(S)] is fixed, we can also write
2 1 yk y var2 (Y!HT ) = − − ∆k . (2.17) 2 πk π k∈U ∈U =k
2.17.3 Variance Estimation From expressions (2.16) and (2.17), it is possible to construct two variance estimators. For general sampling designs, we have: var % 1 (Y!HT ) =
r(Sk )r(S )yk y ∆k . πk π πk
(2.18)
k∈U ∈U
When the reduced sample size n[r(S)] is fixed, we can define the Sen-YatesGrundy estimator (see Sen, 1953; Yates and Grundy, 1953):
2 1 yk y r(Sk )r(S )∆k ! var % 2 (YHT ) = − − . (2.19) 2 πk π πk k∈U ∈U =k
If πk > 0 for all k, ∈ U, var % 1 (Y!HT ) and var % 2 (Y!HT ) are unbiased estimators ! ! of var1 (YHT ) and var2 (YHT ), respectively. Both estimators can take negative values. Nevertheless, if ∆k < 0, for all k = ∈ U, then var % 2 (Y!HT ) ≥ 0. Due to the double sum, both variance estimators are not really applicable. For each particular sampling design, either they can be simplified in a simple sum or an approximation must be used. By developing var % 2 (Y!HT ), we get the following result.
2.18 More on Estimation in Sampling With Replacement
29
Result 10. var % 1 (Y!HT ) = var % 2 (Y!HT ) +
r(Sk )y 2 r(S )∆k k . πk2 πk
k∈U
∈U
Note that, with fixed sample size, estimator (2.18) should never be used.
2.18 More on Estimation in Sampling With Replacement When the sampling design is with replacement, there exist three basic ways to estimate the total: 1. The Hansen-Hurwitz estimator Y!HH =
Sk y k , µk
k∈U
2. The Horvitz-Thompson estimator Y!HT =
r(Sk )yk , πk
k∈U
3. The improved Hansen-Hurwitz estimator E[S |r(S)]y k k . Y!IHH = E Y!HH |r(S) = µk k∈U
These three estimators are unbiased. They are equal when the sampling design is without replacement. From Theorem 4 (page 25) of Rao-Blackwell, we have that var(Y!IHH ) ≤ var(Y!HH ).
(2.20)
The Hansen-Hurwitz estimator is not admissible, in the sense that it is always possible to construct (at least theoretically) an improved estimator that always has a smaller variance than Y!HH . Unfortunately, the conditional expectation E[Sk |r(S)] needed to compute Y!IHH is often very intricate, especially when the units are selected with unequal probabilities.
3 Sampling Algorithms
3.1 Introduction A sampling algorithm is a procedure used to select a sample. If the sampling design is known and if the population size is not too large, a sample can be selected directly by enumerating all the samples as explained in Section 3.3. Nevertheless, when N is large, the number of possible samples becomes so large that it is practically impossible to enumerate all the samples. The objective of a sampling algorithm is to select a sample by avoiding the enumeration of all the samples. The main difficulty of the implementation is thus the combinatory explosion of the number of possible samples. Hedayat and Sinha (1991) and Chen (1998) pointed out that several distinct sampling algorithms (or sampling schemes) can correspond to the same sampling design, which is clearly illustrated in Chapter 4, devoted to simple random sampling. There are, however, several families of algorithms to implement sampling designs. In this chapter, we describe these general algorithms and propose a typology. Indeed, most of the particular algorithms can be presented as particular cases of very general schemes that can be applied to any sampling design.
3.2 Sampling Algorithms We refer to the following definitions: Definition 31. A sampling algorithm is a procedure allowing the selection of a random sample. Definition 32. A sampling algorithm is said to be enumerative if all the possible samples must be listed in order to select the random sample. An efficient sampling algorithm is by definition a fast one. All enumerative algorithms are therefore inefficient. The basic aim is thus to find a shortcut
32
3 Sampling Algorithms
that allows us to select a sample by means of a sampling design and to avoid the complete computation of the sample. It seems that there is no general method for constructing a shortcut for any sampling design. The possibility of constructing a fast algorithm depends on the sampling design that must be implemented. There exist, however, some standard algorithms of sampling that can be applied to any sampling design. Nevertheless, nonenumerative sampling designs are obtained only for very specific designs.
3.3 Enumerative Selection of the Sample In order to implement a sampling design p(.), an enumerative procedure presented in Algorithm 3.1 is theoretically always possible. As presented in TaAlgorithm 3.1 Enumerative algorithm 1. First, construct a list {s1 , s2 , . . . , sj , . . . , sJ } of all possible samples with their probabilities. 2. Next, generate a random variable u with a uniform distribution in [0,1]. j−1 j p(si ) ≤ u < p(si ). 3. Finally, select the sample sj such that i=1
i=1
ble 3.1, even for small population sizes, the size of the support is so large that it becomes unworkable very quickly. A large part of the developments in sampling theory is the research of shortcuts for avoiding this enumeration. Table 3.1. Sizes of symmetric supports SupportQ card(Q) N = 100, n = 10 N = 300, n = 30 R Rn
∞ N +n−1
− 5.1541 × 1013
− 3.8254 × 1042
S
2N N
1.2677 × 1030
2.0370 × 1090
1.7310 × 1013
1.7319 × 1041
Sn
n
n
3.4 Martingale Algorithms A very large family is the set of the martingale algorithms. Almost all sampling methods can be expressed in the form of a martingale algorithm, except the rejective algorithm (on this topic see Section 3.8).
3.5 Sequential Algorithms
33
Definition 33. A sampling algorithm is said to be a martingale algorithm if it can be written as a finite sequence of random vectors µ(0), µ(1), µ(2), . . . , µ(t), . . . , µ(T ), of RN , such that • • • •
µ(0) = µ = E(S), E [µ(t)|µ(t − 1), . . . , µ(0)] = µ(t − 1), µ(t) is in the convex hull of the support, µ(T ) ∈ R.
A martingale algorithm is thus a “random walk” in the convex hull of the support that begins at the inclusion probability vector and stops on a sample. Definition 34. In sampling without replacement, a sampling algorithm is said to be a martingale algorithm if it can be written as a finite sequence of random vectors π(0), π(1), π(2), . . . , π(t), . . . , π(T ), of [0, 1]N , such that • • • •
π(0) = π = E(S), E [π(t)|π(t − 1), . . . , π(0)] = π(t − 1), π(t) is in the convex hull of the support, π(T ) ∈ S.
Note that, in sampling without replacement, the convex hull of the support is a subset of the hypercube [0, 1]N . A martingale algorithm consists thus of modifying randomly at each step the vector of inclusion probabilities until a sample is obtained. All the sequential procedures (see Algorithm 3.2, page 35), the draw by draw algorithms (see Algorithm 3.3, page 35), the algorithms of the splitting family (see Chapter 6), and the cube family of algorithms (see Chapter 8) can be presented as martingales.
3.5 Sequential Algorithms A sequential procedure is a method that is applied to a list of units sorted according to a particular order denoted 1, . . . , k, . . . , N . Definition 35. A sampling procedure is said to be weakly sequential if at step k = 1, . . . , N of the procedure, the decision concerning the number of times that unit k is in the sample is definitively taken. Definition 36. A sampling procedure is said to be strictly sequential if it is weakly sequential and if the decision concerning unit k does not depend on the units that are after k on the list.
34
3 Sampling Algorithms
The standard sequential procedure consists of examining successively all the units of the population. At each one of the N steps, unit k is selected sk times according to a distribution probability that depends on the decision taken for the previous units. The standard sequential procedure presented in Algorithm 3.2 is a weakly sequential method that can be defined and can theoretically be used to implement any sampling design. Algorithm 3.2 Standard sequential procedure 1. Let p(s) be the sampling design and Q the support. First, define p(s), s1 = 0, 1, 2, . . . q1 (s1 ) = Pr(S1 = s1 ) = s∈Q|S1 =s1
2. Select the first unit s1 times according to the distribution q1 (s1 ). 3. For k = 2, . . . , N do a) Compute qk (sk ) = Pr(Sk = sk |Sk−1 = sk−1 , . . . , S1 = s1 ) s∈Q|Sk =sk ,Sk−1 =sk−1 ,...,S1 =s1 p(s) , sk =, 0, 1, 2, . . . = s∈Q|Sk−1 =sk−1 ,...,S1 =s1 p(s) b) Select the kth unit sk times according to the distribution qk (sk ); EndFor.
This procedure provides a nonenumerative procedure only if the qk (sk ) can be computed successively without enumerating all the possible samples of the sampling designs. The standard sequential algorithm is implemented for particular sampling designs in Algorithm 4.1, page 44; Algorithm 4.3, page 48; Algorithm 4.6, page 53; Algorithm 4.8, page 61; Algorithm 5.1, page 70; Algorithm 5.2, page 75; Algorithm 5.4, page 79; and Algorithm 5.8, page 92. Example 4. Suppose that the sampling design is
−1 N p(s) = , s ∈ Sn , n which is actually a simple random sampling without replacement (see Section 4.4). After some algebra, we get ⎧ k−1 n − =1 s ⎪ ⎪ ⎨ if sk = 1 N −k+ qk (sk ) = 1k−1 ⎪ ⎪ ⎩ 1 − n − =1 s if sk = 0. N −k+1 This method was actually proposed by Fan et al. (1962) and Bebbington (1975) and is described in detail in Algorithm 4.3, page 48.
3.6 Draw by Draw Algorithms
35
3.6 Draw by Draw Algorithms The draw by draw algorithms are restricted to designs with fixed sample size. We refer to the following definition. Definition 37. A sampling design of fixed sample size n is said to be draw by draw if, at each one of the n steps of the procedure, a unit is definitively selected in the sample. For any sampling design p(s) of fixed sample size on a support Q ⊂ Rn , there exists a standard way to implement a draw by draw procedure. At each one of the n steps, a unit is selected randomly from the population with probabilities proportional to the mean vector. Next, a new sampling design is computed according to the selected unit. This procedure can be applied to sampling with and without replacement and is presented in Algorithm 3.3. Algorithm 3.3 Standard draw by draw algorithm 1. Let p(s) be a sampling design and Q ⊂ Rn the support. First, define p(0) (s) = p(s) and Q(0) = Q. Define also b(0) as the null vector of RN . 2. For t = 0, . . . , n − 1 do a) Compute ν(t) = s∈Q(t) sp(t) (s); b) Select randomly one unit from U with probabilities qk (t), where νk (t) νk (t) = , k ∈ U; ν (t) n −t ∈U
qk (t) =
The selected unit is denoted j; c) Define aj = (0 · · · 0 1 0 · · · 0); Execute b(t + 1) = b(t) + aj ; j th d) Define Q(t + 1) = {˜s = s − aj , for all s ∈ Q(t) such that sj > 0} ; e) Define, for all ˜s ∈ Q(t + 1),
sj p(t) (s) , (t) (s) s∈Q(t) sj p
p(t+1) (˜s) = where s = ˜s + aj ; EndFor. 3. The selected sample is b(n).
The validity of the method is based on the recursive relation: p(t) (s) =
k∈U
sj p(t) (s) = qk (t)p(t+1) (˜s), (t) s∈Q(t) sj p (s) k∈U
qk (t)
where s ⊂ Q(t), and ˜s = s − aj . As Q(0) ⊂ Rn , we have Q(t) ⊂ Rn−t . Thus,
36
3 Sampling Algorithms
νk (t) = (n − t).
k∈U
Moreover, the sequence µ(t) = ν(t) +
t
b(j),
j=1
for t = 0, . . . , n − 1, is a martingale algorithm. The standard draw by draw procedure can theoretically be implemented for any sampling design but is not necessarily nonenumerative. In order to provide a nonenumerative algorithm, the passage from qk (t) to qk (t+1) should be such that it is not necessary to compute p(t) (s), which depends on the sampling design and on the support. The standard draw by draw algorithm is implemented for particular sampling designs in Algorithm 4.2, page 48; Algorithm 4.7, page 60; Algorithm 5.3, page 76; and Algorithm 5.9, page 95. Example 5. Suppose that the sampling design is p(s) =
n! & 1 , s ∈ Rn , Nn sk ! k∈U
which is actually a simple random sampling with replacement (see Section 4.6). We have, for all t = 0, . . . , n − 1,
ν(t) =
n−t n−t ··· N N
,
1 , k ∈ U, N Q(t) = Rn−t , (n − t)! & 1 p(t) (s) = (n−t) , s ∈ Rn−t . sk ! N qk (t) =
k∈U
It is not even necessary to compute p(t) (s). At each step, a unit is selected from U with probability 1/N for each unit. This method is developed in Algorithm 4.7, page 60. In the case of sampling without replacement, the standard draw by draw algorithm can be simplified as presented in Algorithm 3.4. Example 6. Suppose that the sampling design is
p(s) =
N n
−1 , s ∈ Sn ,
3.7 Eliminatory Algorithms
37
Algorithm 3.4 Standard draw by draw algorithm for sampling without replacement 1. Let p(s) be a sampling design and Q ∈ S the support. 2. Define b = (bk ) = 0 ∈ RN . 3. For t = 0, . . . , n − 1 do select a unit from U with probability 1 E (Sk |Si = 1 for all i such that bi = 1) if bk = 0 qk = n − t 0 if bk = 1; If unit j is selected, then bj = 1; EndFor.
which is actually a simple random sampling without replacement (see Section 4.4). In sampling without replacement, at each step, a unit is definitively selected in the sample. This unit cannot be selected twice. The draw by draw standard method is: n−t if unit k is not yet selected νk (t) = N − t 0 if unit k is already selected, qk (t) =
1 if unit k is not yet selected N −t 0 if unit k is already selected.
Thus, at step t, a unit is selected with equal probabilities 1/(N − t) among the N − t unselected units. The vector π(t) = [πk (t)] = is a martingale, where πk (t) =
1 if unit k is not yet selected N −t 1 if unit k is already selected,
for t = 0, . . . , n − 1. This method is developed in Algorithm 4.2, page 48.
3.7 Eliminatory Algorithms For sampling without replacement and of fixed sample size, it is possible to use methods for which units are eliminated from the population, which can be defined as follows. Definition 38. A sampling algorithm of fixed sample size n is said to be eliminatory if, at each one of the N −n steps of the algorithm, a unit is definitively eliminated from the population.
38
3 Sampling Algorithms
Algorithm 3.5 Standard eliminatory algorithm 1. Let p(s) be the sampling design and Q the support such that Q ⊂ Sn . Define p(0) (s) = p(s) and Q(0) = Q. 2. For t = 0, . . . , N − n − 1 do a) Compute ν(t) = [νk (t)], where p(t) (s) = (1 − sk )p(t) (s); νk (t) = s∈Q(t)|sk =0
s∈Q(t)
b) Select randomly a unit from U with unequal probabilities qk (t), where νk (t) νk (t) = , ν (t) N −n−t ∈U
qk (t) =
k ∈ U;
The selected unit is denoted j; c) Define Q(t + 1) = {s ∈ Q(t)|sj = 0} ; d) Define p(t) (s) , (t) (s) s∈Q(t+1) p
p(t+1) (s) = p(t) (s|Q(t + 1)) =
where s ∈ Q(t + 1); EndFor. 3. The support Q(N − n) contains only one sample that is selected.
Algorithm 3.5 implements a standard eliminatory procedure for a fixed size sampling design without replacement. The eliminatory methods can be presented as the complementary procedures of the draw by draw methods. Indeed, in order to implement an eliminatory method for a sampling design p(.), we can first define the complementary design pc (.) of p(.). Next, a sample s of size N − n is randomly selected by means of a draw by draw method for pc (.). Finally, the sample 1−s is selected, where 1 is a vector of N ones.
3.8 Rejective Algorithms Definition 39. Let Q1 and Q2 be two supports such that Q1 ⊂ Q2 . Also, let p1 (.) be a sampling design on Q1 and p2 (.) be a sampling design on Q2 , such that p2 (s|Q1 ) = p1 (s), for all s ∈ Q1 . A sampling design is said to be rejective if successive independent samples are selected randomly according to p2 (.) in Q2 until a sample of Q1 is obtained. Example 7. Suppose that we want to select a sample with the design
−1 N p1 (s) = , s ∈ Sn , n
3.8 Rejective Algorithms
39
which is a simple random sampling without replacement. Independent samples can be selected with the sampling design p2 (s) =
n! & 1 , s ∈ Rn , Nn sk ! k∈U
which is a simple random sampling with replacement, until a sample of Sn is obtained. Indeed, we have p2 (s|Sn ) = p1 (s),
for all s ∈ Sn .
In practice, rejective algorithms are quite commonly used. The rejective methods are sometimes very fast, and sometimes very slow according to the cardinality difference between Q1 and Q2 . Rejective algorithms are implemented for particular sampling designs in Algorithm 5.5, page 89; Algorithm 5.6, page 90; Algorithm 7.3, page 136; Algorithm 7.4, page 136; and Algorithm 7.5, page 136.
4 Simple Random Sampling
4.1 Introduction In this chapter, a unified theory of simple random sampling is presented. First, an original definition of a simple design is proposed. All the particular simple designs as simple random sampling with and without replacement and Bernoulli sampling with and without replacement can then be deduced by a simple change of symmetric support. For each of these sampling designs, the standard algorithms presented in Chapter 3 are applied, which allows defining eight sampling procedures. It is interesting to note that these algorithms were originally published in statistics and computer science journals without many links between the publications of these fields of research. In sampling with replacement with fixed sample size, the question of estimation is largely developed. It is indeed preferable to suppress the information about the multiplicity of the units, which amounts to applying a Rao-Blackwellization on the Hansen-Hurwitz estimator. The interest of each estimator is thus discussed.
4.2 Definition of Simple Random Sampling Curiously, a concept as common as simple random sampling is often not defined. We refer to the following definition. Definition 40. A sampling design pSIMPLE (., θ, Q) of parameter θ ∈ R∗+ on a support Q is said to be simple, if (i) Its sampling design can be written ' θn(s) k∈U 1/sk ! ' , pSIMPLE (s, θ, Q) = n(s) s∈Q θ k∈U 1/sk !
for all s ∈ Q.
(ii)Its support Q is symmetric (see Definition 2, page 9).
42
4 Simple Random Sampling
Remark 7. If a simple sampling design is defined on a support Q without replacement, then it can be written: θn(s) , n(s) s∈Q θ
pSIMPLE (s, θ, Q) = for all s ∈ Q.
Remark 8. If a simple sampling design has a fixed sample size, then it does not depend on the parameter anymore: ' 1/sk ! ' pSIMPLE (s, θ, Q) = k∈U , s∈Q k∈U 1/sk ! for all s ∈ Q. Remark 9. If a simple sampling design is without replacement and of fixed sample size, then: 1 , pSIMPLE (s, θ, Q) = cardQ for all s ∈ Q. Remark 10. For any s∗ obtained by permuting the elements of s, we have pSIMPLE (s|Q) = pSIMPLE (s∗ |Q). Remark 11. If S is a random sample selected by means of a simple random sampling on Q, then, because the support is symmetric, 1. µ = E(Sk ) and π = E[r(Sk )] do not depend on k, E[n(S)] 2. µ = , N E {n[r(S)]} . 3. π = N The most studied simple sampling designs are: • Bernoulli sampling design (BERN) on S, • Simple random sampling without replacement (with fixed sample size) (SRSWOR) on Sn , • Bernoulli sampling with replacement (BERNWR) on R, • Simple random sampling with replacement with fixed sample size (SRSWR) on Rn .
4.3 Bernoulli Sampling (BERN)
43
4.3 Bernoulli Sampling (BERN) 4.3.1 Sampling Design Definition 41. A simple design defined with parameter θ on support S = {0, 1}N is called a Bernoulli sampling design with inclusion probabilities π = θ/(1 + θ). The Bernoulli sampling design can be deduced from Definition 40, page 41. Indeed,
θ pBERN s, π = 1+θ θn(s) θn(s) θn(s) = N = n(s) N r N s∈S θ θr s∈Sr θ r=0 r=0
= pSIMPLE (s, θ, S) = n(s)
=
θ = (θ + 1)N
θ 1+θ
n(s)
1−
θ 1+θ
N −n(s)
r
= π n(s) (1 − π)N −n(s) ,
for all s ∈ S, where π = θ/(1 + θ), π ∈]0, 1[. Bernoulli sampling design can also be written & pBERN (s, π) = π sk (1 − π)1−sk , k∈U
where the S1 , . . . , SN are independent and identically distributed Bernoulli variables with parameter π. The characteristic function is φBERN (t) = eit s pBERN (s, π) = eit s π n(s) (1 − π)N −n(s) s∈S
s∈S
sk & π &
π itk 1+ = (1 − π)N = (1 − π)N eitk e 1−π 1−π s∈S k∈U k∈U & 1 − π + πeitk . = k∈U
By differentiating the characteristic function, we get the inclusion probabilities E(Sk ) = π, for all k ∈ U. By differentiating the characteristic function twice, we get the joint inclusion probabilities E(Sk S ) = π 2 , for all k = ∈ U. The sample size distribution is binomial, n(S) ∼ B(N, π), and
N Pr [n(s) = r] = π r (1 − π)N −r = π r (1 − π)N −r , r s∈Sr
with r = 0, . . . , N.
44
4 Simple Random Sampling
Because the Sk ’s have a Bernoulli distribution, the variance is var(Sk ) = π(1 − π), for all k ∈ U. Due to the independence of the Sk , ∆k = cov(Sk , S ) = 0, for all k = ∈ U. The variance-covariance operator is Σ = π(1 − π)I, where I is an (N × N ) identity matrix. 4.3.2 Estimation The Horvitz-Thompson estimator is 1 y k Sk . Y!HT = π k∈U
The variance of the Horvitz-Thompson estimator is ˇ ∆ˇ y= var(Y!HT ) = y
π(1 − π) 2 yk . π2 k∈U
The variance estimator of the Horvitz-Thompson estimator is var( % Y!HT ) =
1−π 2 yk Sk . π2 k∈U
4.3.3 Sequential Sampling Procedure for BERN Due to the independence of the sampling between the units, the standard sequential Algorithm 3.2, page 34, is very easy to construct for a BERN sampling as presented in Algorithm 4.1. Algorithm 4.1 Bernoulli sampling without replacement Definition k : Integer; For k = 1, . . . , N do with probability π select unit k; EndFor.
Note that Algorithm 4.1 is strictly sequential. Conditionally on a fixed sample size, a Bernoulli design becomes a simple random sample without replacement. Another method for selecting a sample with a Bernoulli design is thus to choose randomly the sample size according to a binomial distribution and then to select a sample with a simple random sampling without replacement.
4.4 Simple Random Sampling Without Replacement (SRSWOR)
45
4.4 Simple Random Sampling Without Replacement (SRSWOR) 4.4.1 Sampling Design Definition 42. A simple design defined with parameter θ on support Sn is called a simple random sampling without replacement. The simple random sampling without replacement can be deduced from Definition 40, page 41: pSRSWOR (s, n) = pSIMPLE (s, θ, Sn ) = 1 = card(Sn )
=
N n
−1
θn
s∈Sn
=
θn
n!(N − n)! , N!
for all s ∈ Sn . Note that pSRSWOR (s, n) does not depend on θ anymore. The characteristic function is
−1 N eit s pSRSWOR,n (s) = eit s , φSRSWOR (t) = n s∈Sn
s∈Sn
and cannot be simplified. The inclusion probability is πk =
sk pSRSWOR (s, n) = card{s ∈ Sn |sk = 1}
s∈Sn
=
N −1 n−1
N n
−1 =
N n
−1
n , N
for all k ∈ U. The joint inclusion probability is πk =
sk s pSRSWOR (s, n) = card{s ∈ Sn |sk = 1 and s = 1}
s∈Sn
=
N −2 n−2
N n
−1 =
N n
−1
n(n − 1) , N (N − 1)
for all k = ∈ U. The variance-covariance operator is given by ⎧ n(N − n) ⎪ ⎨ πk (1 − πk ) = , N2 ∆k = cov(Sk , S ) = 2 n n(N − n) n(n − 1) ⎪ ⎩ − 2 =− 2 , N (N − 1) N (N − 1) N
k=∈U k = ∈ U.
46
4 Simple Random Sampling
Thus,
n(N − n) H, N (N − 1) where H is the projection matrix that centers the data. ⎞ ⎛ 1 − N1 · · · −1 · · · −1 N N ⎜ .. .. .. ⎟ .. ⎜ . . . . ⎟ ⎟ ⎜ −1 11 1 −1 ⎟ ⎜ ,= ⎜ N ··· 1 − N ··· N ⎟, H=I− N ⎜ . .. .. ⎟ .. ⎝ .. . . . ⎠ −1 −1 1 · · · · · · 1 − N N N ∆ = var(S) =
(4.1)
and I is an identity matrix and 1 is a vector of N ones. We have N Hˇ y= y1 − Y · · · yk − Y · · · yN − Y , n ˇ= where y
N n
(y1 · · · yk · · · yN ) . Thus, ˇ Hˇ y= y
2 N2 yk − Y . n2 k∈U
In order to estimate the variance, we need to compute ⎧ ⎪n−1 ∆k n(N − n) ⎨ n , k = ∈ U = πk N (n − 1) ⎪ ⎩−1, k = ∈ U. n 4.4.2 Estimation Horvitz-Thompson estimator The Horvitz-Thompson estimator is N Y!HT = y k Sk . n k∈U
Variance of the Horvitz-Thompson estimator The variance of the Horvitz-Thompson estimator is ˇ ∆ˇ var(Y!HT ) = y y=
2 n(N − n) n(N − n) N 2 ˇ Hˇ yk − Y y= y N (N − 1) N (N − 1) n2 k∈U
2 1 (N − n) N − n Vy2 yk − Y = N 2 = N2 , N n(N − 1) N n k∈U
where Vy2 =
2 1 yk − Y¯ . N −1 k∈U
4.4 Simple Random Sampling Without Replacement (SRSWOR)
47
Variance estimator of the Horvitz-Thompson estimator The variance estimator of the Horvitz-Thompson estimator is N −n 2 v , nN y
var( % Y!HT ) = N 2 where
1 = Sk n−1
vy2
(
Y!HT yk − N
k∈U
(4.2) )2 .
Note that (4.2) is a particular case of estimators (2.18) and (2.19), page 28, that are equal in this case. Estimator (4.2) is thus unbiased, but a very simple proof can be given for this particular case: Result 11. In a SRSWOR vy2 is an unbiased estimator of Vy2 . Proof. Because vy2
1 = Sk n−1
(
k∈U
Y!HT yk − N
)2 =
1 2 (yk − y ) Sk S , 2n(n − 1) k∈U ∈U
we obtain 1 2 (yk − y ) E (Sk S ) 2n(n − 1) k∈U ∈U 1 2 = (yk − y ) πk 2n(n − 1)
E(vy2 ) =
k∈U ∈U
1 2 n(n − 1) (yk − y ) 2n(n − 1) N (N − 1) k∈U ∈U 1 2 (yk − y ) = Vy2 . = 2N (N − 1)
=
2
k∈U ∈U
4.4.3 Draw by Draw Procedure for SRSWOR The standard draw by draw Algorithm 3.3, page 35, becomes in the case of SRSWOR the well-known ball-in-urn method (see Example 4, page 34) where a unit is selected from the population with equal probability. The selected unit is removed from the population. Next, a second unit is selected among the N − 1 remaining units. This operation is repeated n times and, when a unit is selected, it is definitively removed from the population. This method is summarized in Algorithm 4.2. Nevertheless, Algorithm 4.2 has an important drawback: its implementation is quite complex and it is not sequential. For this reason, it is preferable to use the selection-rejection procedure presented in Section 4.4.4.
48
4 Simple Random Sampling
Algorithm 4.2 Draw by draw procedure for SRSWOR Definition j : Integer; For t = 0, . . . , n − 1 do select a unit
1k from the population with probability if k is not already selected qk = N −t 0 if k is already selected; EndFor.
4.4.4 Sequential Procedure for SRSWOR: The Selection-Rejection Method The selection-rejection procedure is the standard sequential Algorithm 3.2, page 34, for SRSWOR; that is, the sample can be selected in one pass in the data file. The method was discussed in Fan et al. (1962), Bebbington (1975), Vitter (1984), Ahrens and Dieter (1985) and Vitter (1987) and is presented in Algorithm 4.3. The selection-rejection method is the best algorithm for selecting a sample according to SRSWOR; it is strictly sequential. Note that the population size must be known before applying Algorithm 4.3. Algorithm 4.3 Selection-rejection procedure for SRSWOR Definition k, j : Integer; j = 0; For k = 1, . . . , N do select unit k; n−j with probability then N − (k − 1) j = j + 1; EndFor.
This method was improved by Fan et al. (1962), Knuth (1981), Vitter (1987), Ahrens and Dieter (1985), Bissell (1986), and Pinkham (1987) in order to skip the nonselected units directly. These methods are discussed in Deville et al. (1988). 4.4.5 Sample Reservoir Method The reservoir method has been proposed by McLeod and Bellhouse (1983) and Vitter (1985) but is also a particular case of the Chao procedure (see Chao, 1982, and Section 6.3.6, page 119). As presented in Algorithm 4.4, at the first step, the first n units are selected in the sample. Next, the sample is updated by examining the last N − n units of the file. The main interest of Algorithm 4.4 is that the population size must not be known before applying the algorithm. At the end of the file, the population size is known. The reservoir method can be viewed as an eliminatory algorithm. Indeed, at each one of the N − n steps of the algorithm, a unit is definitively eliminated.
4.4 Simple Random Sampling Without Replacement (SRSWOR)
49
Algorithm 4.4 Reservoir method for SRSWOR Definition k : Integer; The first n units are selected into the sample; For k = n + 1, . . . , N do select unit k; with probability n a unit is removed from the sample with equal probabilities; k unit k takes the place of the removed unit; EndFor.
Result 12. (McLeod and Bellhouse, 1983) Algorithm 4.4 gives a SRSWOR. Proof. (by induction) Let Uj denote the population of the first j units of U with j = n, n + 1, . . . , N, and Sn (Uj ) = {s ∈ Sn |sk = 0, for k = j + 1, . . . , N } . Let S(j) denote a random sample of size n selected in support Sn (Uj ). All the vectors s(j) = (s1 (j) · · · sk (j) · · · sN (j)) of Sn (Uj ) are possible values for the random sample S(j). We prove that Pr [S(j) = s(j)] =
−1 j , for s(j) ⊂ Sn (Uj ). n
(4.3)
Expression (4.3) is true when j = n. Indeed, in this case, Sn (Uj ) only contains a sample s(n) = Un and Pr(Sn = sn ) = 1. When j > n, Expression (4.3) is proved by induction, supposing that (4.3) is true for j − 1. Considering a sample s(j) of Sn (Uj ), two cases must be distinguished. • Case 1. sj (j) = 0 We obtain Pr[S(j) = s(j)] = Pr[S(j − 1) = s(j)] × Pr(unit j is not selected at step j)
−1
−1 j−1 j n = = 1− . n j n
(4.4)
• Case 2. sj (j) = 1 In this case, the probability of selecting sample s(j) at step j is the probability that: 1. At step j − 1, a sample s(j) − aj + ai is selected, where i is any unit of Uj−1 that is not selected in sj and aj = (0 · · · 0 *+,1 0 · · · 0); j th
2. Unit j is selected; 3. Unit i is replaced by j.
50
4 Simple Random Sampling
This probability can thus be written 1 n Pr[Sj = s(j)] = Pr [S(j − 1) = s(j) − aj + ai ] × × j n i∈Uj−1 |si (j)=0
−1
−1 j−1 j 1 = (j − n) = . (4.5) n j n Results (4.4) and (4.5) prove Expression (4.3) by induction. Finally, Result 12 is obtained by taking j = N in Expression (4.3). 2 4.4.6 Random Sorting Method for SRSWOR A very simple method presented in Algorithm 4.5 consists of randomly sorting the population file. Algorithm 4.5 Random sort procedure for SRSWOR 1. A value of an independent uniform variable in [0,1] is allocated to each unit of the population. 2. The population is sorted in ascending (or descending) order. 3. The first (or last) n units of the sorted population are selected in the sample.
Sunter (1977, p. 273) has proved that the random sorting gives a SRSWOR. Result 13. The random sorting method gives a SRSWOR. Proof. Consider the distribution function of the largest uniform random number generated for N − n units. If ui denotes the uniform random variable of the ith unit of the file, this distribution function is given by: N −n N −n & F (x) = Pr (ui ≤ x) = Pr (ui ≤ x) = xN −n , 0 < x < 1. i=1
i=1
The probability that the n remaining units are all greater than x is equal to (1 − x)n . The probability of selecting a particular sample s is thus . 1 . 1 n p(s) = (1 − x) dF (x) = (N − n) (1 − x)n xN −n−1 dx, (4.6) 0
0
which is an Euler integral . 1 tx (1 − t)y dt = B(x, y) = 0
x!y! , (x + y + 1)!
x, y integer.
(4.7)
From (4.6) and (4.7), we get
p(s) =
N n
−1 .
2
4.5 Bernoulli Sampling With Replacement (BERNWR)
51
4.5 Bernoulli Sampling With Replacement (BERNWR) 4.5.1 Sampling Design Bernoulli sampling with replacement is not useful in practice. However, we introduce it for the completeness of the definitions. Definition 43. A simple design defined on support R is called Bernoulli sampling design with replacement. The Bernoulli sampling design with replacement can be deduced from Definition 40, page 41: ' µn(s) k∈U s1k ! ' pBERNWR (s, µ) = pSIMPLE (s, θ = µ, R) = 1 n(s) s∈R µ k∈U sk ! ' & 1 µn(s) k∈U s1k ! −N µ n(s) , = ' µ ∞ µi = e sk ! i=0 k∈U k∈U
i!
for all s ∈ R and with µ ∈ R∗+ . Because we can also write the sampling design as pBERNWR (s, µ) =
& e−µ µsk , sk !
k∈U
where the S1 , . . . , SN are independent and identically distributed Poisson variables with parameter µ, the following results can be derived directly from the Poisson distribution. The expectation is µk = E(Sk ) = µ, for all k ∈ U. The joint expectation is µk = E(Sk S ) = E(Sk )E(S ) = µ2 , for all k = ∈ U. In a Poisson distribution, the variance is equal to the expectation Σkk = var(Sk ) = µ, for all k ∈ U. Due to the independence of the Sk , Σk = cov(Sk , S ) = 0, for all k = ∈ U. The variance-covariance operator is Σ = Iµ. Because the sum of independent Poisson variables is a Poisson variable, the probability distribution of the sample size comes directly: Pr [n(S) = z] =
e−(N µ) (N µ)z , z!
z ∈ N.
52
4 Simple Random Sampling
The characteristic function is & e−µ µsk sk ! s∈R s∈R k∈U it sk it sk ∞ & e kµ & e kµ = e−N µ = e−N µ sk ! sk ! s∈R k∈U k∈U sk =0 & & / 0 exp eitk µ = exp µ eitk − 1 = exp µ (exp itk − 1). = e−N µ
φBERNWR (t) =
eit s pBERNWR (s, µ) =
k∈U
eit s
k∈U
k∈U
The inclusion probability is πk = Pr(Sk > 0) = 1 − Pr(Sk = 0) = 1 − e−µ . The joint inclusion probability is πk = Pr(Sk > 0 and S > 0) = [1 − Pr(Sk = 0)] [1 − Pr(S = 0)] = (1 − e−µ )2 ,
k = .
Finally, we have ∆ = I(1 − e−µ )e−µ , where I is an N × N identity matrix. 4.5.2 Estimation Hansen-Hurwitz estimator The Hansen-Hurwitz estimator is Y!HH =
y k Sk . µ
k∈U
The variance of the Hansen-Hurwitz estimator is var(Y!HH ) =
y2 k . µ
k∈U
The estimator of the variance is var( % Y!HH ) =
Sk y 2 k . µ2
k∈U
Improvement of the Hansen-Hurwitz estimator Because E(Sk |r(S)) = r(Sk )
µ , 1 − e−µ
4.6 Simple Random Sampling With Replacement (SRSWR)
53
the improved Hansen-Hurwitz estimator is Y!IHH =
yk E(Sk |r(S)) yk r(Sk ) yk r(Sk ) µ = . = µ µ 1 − e−µ 1 − e−µ
k∈U
k∈U
−µ
k∈U
−µ
As var[r(Sk )] = (1 − e )e , the variance of the improved Hansen-Hurwitz estimator is given by y 2 var[r(S )] (1 − e−µ )e−µ y2 k k k var Y!IHH = . = yk2 = −µ 2 −µ 2 (1 − e ) (1 − e ) eµ − 1 k∈U
k∈U
k∈U
The improvement brings an important decrease of the variance with respect to the Hansen-Hurwitz estimator. This variance can be estimated by r(S )y 2 k k var % Y!IHH = . (eµ − 1)2 k∈U
The Horvitz-Thompson estimator Because πk = 1 − e−µ , the Horvitz-Thompson estimator is Y!HT =
yk r(Sk ) yk r(Sk ) = , πk 1 − e−µ
k∈U
k∈U
which is, in this case, the same estimator as the improved Hansen-Hurwitz estimator. 4.5.3 Sequential Procedure for BERNWR Due to the independence of the Sk , the standard sequential Algorithm 3.2, page 34, is strictly sequential and simple to implement for a BERNWR, as presented in Algorithm 4.6. Algorithm 4.6 Bernoulli sampling with replacement Definition k : Integer; For k = 1, . . . , N do select randomly sk times unit k according to the Poisson distribution P(µ); EndFor.
4.6 Simple Random Sampling With Replacement (SRSWR) 4.6.1 Sampling Design Definition 44. A simple design defined on support Rn is called a simple random sampling with replacement.
54
4 Simple Random Sampling
The simple random sampling with replacement can be deduced from Definition 40, page 41: ' θn(s) k∈U s1k ! ' pSRSWR (s, n) = pSIMPLE (s, θ, Rn ) = 1 n(s) s∈Rn θ k∈U sk !
& 1 sk n! n! & 1 = = n , N sk ! s1 ! . . . sN ! N k∈U
k∈U
for all s ∈ Rn . Note that pSRSWR (s, n) does not depend on θ anymore. The sampling design is multinomial. The characteristic function is n! & 1 eit s pSRSWR (s, n) = eit s n φSRSWR (t) = N sk ! s∈Rn s∈Rn k∈U ( )n
& eitk sk 1 1 = = n! exp itk . N sk ! N s∈Rn
The expectation of S is µ =
k∈U
n
n
k∈U
· · · N . The joint expectation is ⎧ n(N − 1 + n) ⎪ ⎨ , k= N2 µk = E(Sk S ) = ⎪ ⎩ n(n − 1) , k = . N2 The variance-covariance operator is ⎧ n(N − 1) n(N − 1 + n) n2 ⎪ ⎨ , k= − 2 = N N N2 Σk = E(Sk S ) = 2 ⎪ ⎩ n(n − 1) − n = − n , k = . N2 N2 N2 Thus Σ = var(S) = Hn/N, where H is defined in (4.1). Moreover, ⎧ (N − 1) ⎪ ⎨ , k= Σk = (N −11 + n) ⎪ µk ⎩− , k = . (n − 1) N
The inclusion probability is πk = Pr(Sk > 0) = 1 − Pr(Sk = 0) = 1 −
N −1 N
n .
The joint inclusion probability is πk = Pr(Sk > 0 and S > 0) = 1 − Pr(Sk = 0 or S = 0) = 1 − Pr(Sk = 0) − Pr(S = 0) + Pr(Sk = 0 and S = 0) n
n
N −2 N −1 + . = 1−2 N N
(4.8)
4.6 Simple Random Sampling With Replacement (SRSWR)
55
4.6.2 Distribution of n[r(S)] The question of the distribution of n[r(S)] in simple random sampling with replacement has been studied by Basu (1958), Raj and Khamis (1958), Pathak (1962), Pathak (1988), Chikkagoudar (1966), Konijn (1973, chapter IV), and Cassel et al. (1993, p. 41). In sampling with replacement, it is always preferable to suppress the information about the multiplicity of the units (see Theorem 3, page 24). If we conserve only the distinct units, the sample size becomes random. The first objective is thus to compute the sample size of the distinct units. Result 14. If n[r(S)] is the number of distinct units obtained by simple random sampling of m units with replacement in a population of size N , then Pr {n[r(S)] = z} = where
N! s(z) , (N − z)!N m m
z = 1, . . . , min(m, N ),
(4.9)
s(z) m is a Stirling number of the second kind s(z) m =
1 z m i (−1)z−i . z! i=1 i z
Proof. (by induction) The Stirling numbers of the second kind satisfy the recursive relation: (z−1) (z) s(z) (4.10) m = sm−1 + z sm−1 , with the initial values s1 = 1 and s1 = 0, z = 1 (see Abramowitz and Stegun, 1964, pp. 824-825). If n[r(Sm )] is the number of distinct units obtained by simple random sampling with replacement of m − 1 units in a population of size N , then we have: (1)
(z)
Pr {n[r(Sm )] = z} = Pr {n[r(Sm−1 )] = z − 1} × Pr {select at the mth draw a unit not yet selected given that n[r(Sm−1 )] = z − 1} + Pr {n[r(Sm−1 )] = z} × Pr {select at the mth draw an unit already selected given that n[r(Sm−1 )] = z} N −z+1 z = Pr {n[r(Sm−1 )] = z − 1} + Pr {n[r(Sm−1 )] = z} . N N (4.11) Moreover, we have the initial conditions, Pr {n[r(S1 )] = z} =
1 if z = 1 0 if z = 1.
56
4 Simple Random Sampling
If we suppose that Result 14 is true at step m − 1, and by property (4.10), the recurrence equation is satisfied by Expression (4.9). Indeed N −z+1 z + Pr {n[r(Sm−1 )] = z} N N N! N −z+1 z N! + s(z−1) s(z) = (N − z + 1)!N m−1 m−1 N (N − z)!N m−1 m−1 N N! N! (z−1) (z) s + z s s(z) = Pr {n[r(Sm )] = z} . = m−1 m−1 = m (N − z)!N (N − z)!N m m
Pr {n[r(Sm−1 )] = z − 1]}
2
As the initial conditions are also satisfied, Result 14 is proved. The following result is used to derive the mean and variance of n[r(S)].
Result 15. If n[r(S)] is the number of distinct units by selecting m units with replacement in a population of size N , then j−1 & (N − j)m N ! E (N − n[r(S)] − i) = m , j = 1, . . . , N. (4.12) N (N − j)! i=0 Proof. E
(j−1 &
) {N − n[r(S)] − i}
i=0
j−1 &
z=1
i=0
min(m,N )
=
min(m,N −j)
=
z=1
=
N! s(z) (N − z)!N m m
N! (N − z)! s(z) (N − z − j)! (N − z)!N m m
(N − j)m N ! N m (N − j)!
Because
(N − z − i)
min(m,N −j)
z=1
(N − j)! s(z) . (N − z − j)!(N − j)m m
(4.13)
(N − j)! s(z) (N − z − j)!(N − j)m m
is the probability of obtaining exactly z distinct units by selecting m units in a population of size N − j, we obtain min(m,N −j)
z=1
(N − j)! s(z) = 1, (N − z − j)!(N − j)m m
and by (4.13), we get directly (4.12).
2
4.6 Simple Random Sampling With Replacement (SRSWR)
By (4.12) when j = 1, the expectation can be derived m 2 1
N −1 E {n[r(S)]} = N 1 − . N
57
(4.14)
By (4.12) when j = 2, after some algebra, the variance can be obtained var {n[r(S)]} =
(N − 1)m (N − 2)m (N − 1)2m + (N − 1) − . m−1 m−1 N N N 2m−2
(4.15)
4.6.3 Estimation Hansen-Hurwitz estimator The Hansen-Hurwitz estimator is
N y k Sk , Y!HH = n k∈U
and its variance given in (2.12) and (2.13) is ( )2 ( )2 2 1 1 1 N 1 var(Y!HH ) = yk − yk2 − yk = yk n n n N N k∈U
k∈U
k∈U
k∈U
N 2 σy2 , = n where σy2 =
(4.16)
2 1 1 2 yk − Y¯ = (yk − y ) . N 2N 2 k∈U
k∈U ∈U
For estimating the variance, we have the following result. Result 16. In a SRSWR, vy2 is an unbiased estimator of σy2 , where )2 ( 1 Y!HH 2 Sk yk − . vy = n−1 N k∈U
Proof. Because vy2
1 = Sk n−1 k∈U
we obtain E(vy2 ) =
(
Y!HH yk − N
)2 =
1 2 (yk − y ) Sk S , 2n(n − 1) k∈U ∈U
1 2 (yk − y ) E (Sk S ) 2n(n − 1) k∈U ∈U
1 2 n(n − 1) (yk − y ) = 2n(n − 1) N2 k∈U ∈U 1 2 = (yk − y ) = σy2 . 2N 2 k∈U ∈U
2
58
4 Simple Random Sampling
Note that depending on the sampling design being a SRSWR or a SRSWOR, the sample variance vy2 estimates distinct functions of interest (see Result 16, page 57 and Result 11, page 47). By Result 16 and Expression (4.16), we have an unbiased estimator of the variance: N 2 vy2 var % 2 (Y!HH ) = . (4.17) n Estimator (4.17) is also a particular case of the estimator given in Expression (2.15), page 27. The other unbiased estimator given in Expression (2.14), page 27, provides a very strange expression. Indeed, by Result 9, page 27 and Expression (4.8), page 54, we have Sk y 2 S Σk k % 2 (Y!HH ) + var % 1 (Y!HH ) = var µ2k µk k∈U ∈U 2 1 N 2 vy2 yk2 Nn n 2 + − Sk . S = n µ2k k (n − 1)(N − 1 + n) n−1 k∈U
Although unbiased, var % 1 (Y!HH ) should never be used. Improved Hansen-Hurwitz estimator Because E [Sk |r(S)] = r(Sk )
n , n[r(S)]
the improved Hansen-Hurwitz estimator is N N Y!IHH = yk E [Sk |r(S)] = yk r(Sk ), n n[r(S)] k∈U
and its variance is
k∈U
N var(Y!IHH ) = E var yk r(Sk ) n[r(S)] n[r(S)] k∈U N + var E yk r(Sk ) n[r(S)] . n[r(S)]
k∈U
The improved Hansen-Hurwitz estimator is unbiased conditionally on n[r(S)], thus the second term vanishes. Conditionally on n[r(S)], the sample is simple with fixed sample size, thus N ! yk r(Sk ) n[r(S)] var(YIHH ) = E var n[r(S)] k∈U "
Vy2 1 1 2 N − n[r(S)] 2 − Vy2 . =N E =E N N n[r(S)] N n[r(S)]
4.6 Simple Random Sampling With Replacement (SRSWR)
59
In order to derive the variance, we have to compute E [1/n(S)] . Result 17. If n[r(S)] is the number of distinct units obtained by selecting m units according to a SRSWR, then E
1 n[r(S)]
" =
N 1 m−1 j . N m j=1
(4.18)
Proof. (by induction) Let n[r(S), N ] be the number of distinct units obtained by selecting m units according to a SRSWR from a population of size N . We first show that " " (N − 1)m Nm −E = N m−1 . E (4.19) n[r(S), N ] n[r(S), N − 1] Indeed, 1 E
(N − 1)m Nm − n[r(S), N ] n[r(S), N − 1]
min(m,N )
=
z=1
1 N! s(z) − z (N − z)! m
min(m,N )
= N m−1
z=1
2
min(m,N −1)
z=1
1 (N − 1)! s(z) z (N − 1 − z)! m
N! s(z) = N m−1 . (N − z)!N m m
By (4.19), we get a recurrence equation " " 1 (N − 1)m 1 1 = + . E E n[r(S), N ] N Nm n[r(S), N − 1] The initial condition is obvious: E
1 n[r(S), 1]
" = 1.
Expression (4.18) satisfies the recurrence equation and the initial condition. 2 The variance is thus: ⎛
⎞ −1 N Vy2 N 1 1 var(Y!IHH ) = N 2 ⎝ m j m−1 − ⎠ Vy2 = m−2 j m−1 . N j=1 N N j=1
60
4 Simple Random Sampling
The Horvitz-Thompson estimator The Horvitz-Thompson estimator is n 2−1 1
N Y!HT = 1 − yk r(Sk ). N −1 k∈U
The variance is thus var Y!HT # $ # $ = E var Y!HT |n[r(S)] + var E Y!HT |n[r(S)]
# $ # $ 1 !IHH |n[r(S)] + 1 var E n[r(S)]Y!IHH |n[r(S)] E var n[r(S)] Y π2 π2 # $ Y 2 1 2 = 2 E {n[r(S)]} var Y!IHH |n[r(S)] + 2 var {n[r(S)]} π ( )π 2 Y2 N − n[r(S)] Vy 1 2 + 2 var {n[r(S)]} = 2 E {n[r(S)]} N 2 π N n[r(S)] π =
Y2 N Vy2 2 + 2 var [n[r(S)]] E N n[r(S)] − n[r(S)] π2 π Y2
N Vy2 2 N E {n[r(S)]} − var n[r(S)]2 + (E {n[r(S)]}) + 2 var {n[r(S)]} . = 2 π π =
This variance is quite complex. Estimator Y!IHH does not necessarily have a smaller variance than Y!HT . However, Y!HT should not be used because it depends directly on Y 2 ; that means that even if the yk ’s are constant, var(Y!HT ) is not equal to zero. We thus advocate for the use of Y!IHH . 4.6.4 Draw by Draw Procedure for SRSWR The standard draw by draw Algorithm 3.3, page 35, gives a well-known procedure for simple random sampling: at each of the N steps of the algorithm, a unit is selected from the population with equal probability (see Example 5, page 36). This method is summarized in Algorithm 4.7. Algorithm 4.7 Draw by draw procedure for SRSWR Definition j : Integer; For j = 1, . . . , n do a unit is selected with equal probability 1/N from the population U ; EndFor.
4.7 Links Between the Simple Sampling Designs
61
4.6.5 Sequential Procedure for SRSWR As SRSWOR has a fixed sample size, a standard sequential Algorithm 3.2, page 34, can be easily implemented. This neglected method, presented in Algorithm 4.8, is better than the previous one because it allows selecting the sample in one pass. Algorithm 4.8 Sequential procedure for SRSWR Definition k, j : Integer; j = 0; For k = 1, . . . , N do select the kth unit according to the binomial distribution sk times k−1 1 si , ; B n− N − k+1 i=1 EndFor.
4.7 Links Between the Simple Sampling Designs The following relations of conditioning can be proved using conditioning with respect to a sampling design: ∗ • pBERNWR (s, µ|Rn ) = pSRSWR µ ∈ R+ ,
(s, n), for all θ , for all θ ∈ R∗+ , • pBERNWR (s, θ|S) = pBERN s, π = 1+θ • pBERNWR (s, θ|Sn ) = pSRSWOR (s, n), for all θ ∈ R∗+ , • pSRSWR (s, n|Sn ) = pSRSWOR (s, n), • pBERN (s, π|Sn ) = pSRSWOR (s, n), for all π ∈]0, 1[.
These relations are summarized in Figure 4.1 and Table 4.1, page 62.
reduction
BERN
conditioning on S
conditioning on Sn conditioning on Sn
BERNWR conditioning on Rn with 0 ≤ n ≤ N
SRSWOR conditioning on Sn
SRSWR Fig. 4.1. Links between the main simple sampling designs
4 Simple Random Sampling 62
Design Q
p(s)
Notation
Support
SRSWR
BERN
Table 4.1. Main simple random sampling designs
BERNWR
π n(s) (1 − π)N −n(s)
Expectation
Sample size
Replacement
πk
µk
n(S)
WOR/WR
µ
1 − e−µ
µ
random
with repl.
2 y k µ k∈U y2 k Sk µ2 k∈U
N2
vy2 n
σy2 n
k∈U
N2
fixed n N n N −1 1− N n(N − 1) N2 n(n − 1) N2 n − 2 N N yk Sk n
with repl.
k∈U
−1
SRSWOR
N n
without repl.
without repl.
n
random π π
π(1 − π) π2 0
yk2 Sk
N2
N2
N − n vy2 N n
N − n Vy2 N n
k∈U
fixed n N n N n(N − n) N2 n(n − 1) N (N − 1) n(N − n) − 2 N (N − 1) N yk Sk n
k∈U
k∈U
yk Sk π k∈U 1−π 2 yk π (1 − π)
k∈U
Rn S Sn
n N −1 1 − 1) eitk eit s 1 + π eitk − 1 N n s∈S
k∈U
n! 1 Nn sk !
R
k∈U
µn(s) 1 eN µ sk !
Inclusion probability Σkk
µ2
(eitk
Variance
µk , k =
exp µ
Joint expectation
Σk , k =
0 yk Sk µ
φ(t)
Covariance
Y
Char. function
Basic est.
var(Y )
k∈U
Variance
var( Y )
k∈U
Variance est.
5 Unequal Probability Exponential Designs
5.1 Introduction A sampling design is a multivariate discrete distribution and an exponential design is thus an exponential multivariate discrete distribution. Exponential designs are a large family that includes simple random sampling, multinomial sampling, Poisson sampling, unequal probability sampling with replacement, and conditional or rejective Poisson sampling. A large part of the posthumous book of H´ ajek (1981) is dedicated to the exponential family. H´ ajek advocated for the use of Poisson rejective sampling but the link between the inclusion probabilities of Poisson sampling and conditional Poisson sampling was not yet clearly elucidated. Chen et al. (1994) have taken a big step forward by linking the question of conditional Poisson sampling to the general theory of the exponential family. The fundamental point developed by Chen et al. (1994) is the link between the parameter of the exponential family and its mean; that is, its vector of inclusion probabilities. Once this link is clarified, the implementation of the classical algorithms follows quite easily. Independently, Jonasson and Nerman (1996), Aires (1999, 2000a), Bondesson et al. (2004), and Traat et al. (2004) have investigated the question of conditional Poisson sampling. Chen and Liu (1997), Chen (1998), and Deville (2000) have improved the algorithms and the technique of computation of inclusion probabilities. A large part of the material of this chapter has been developed during informal conversations with Jean-Claude Deville. In this chapter, we attempt to present a coherent theory of exponential designs. A unique definition is given, and all exponential designs can be derived by changes of support. The application of the classical algorithms presented in Chapter 3 allows for the deduction of nine sampling algorithms.
64
5 Unequal Probability Exponential Designs
5.2 General Exponential Designs 5.2.1 Minimum Kullback-Leibler Divergence In order to identify a unique sampling design that satisfies a predetermined vector of means µ, a general idea is to minimize the Kullback-Leibler divergence (see Kullback, 1959): H(p, pr ) =
p(s) log
s∈Q
p(s) , pr (s)
where pr (s) is a design of reference on Q. The divergence H(p, pr ) (also called relative entropy) is always positive and H(p, pr ) = 0 when p(s) = pr (s), for all s ∈ Q. The objective is thus to identify the closest sampling design (in the sense of the Kullback-Leibler divergence) to pr (s) that satisfies fixed inclusion probabilities. In this chapter, the study of minimum Kullback-Leibler divergence designs is restricted to the case where the reference sampling design is simple; that is, ' θn(s) k∈U s1k ! ' pr (s) = pSIMPLE (s, θ, Q) = (5.1) 1 . n(s) s∈Q θ k∈U sk ! Because H(p, pr ) is strictly convex with respect to p, the minimization of H(p, pr ) under linear constraints provides a unique solution. The KullbackLeibler divergence H(p, pr ) is minimized under the constraints p(s) = 1, (5.2) s∈Q
and
sp(s) = µ.
(5.3)
s∈Q ◦
The vector µ belongs to Q (the interior of Q) (see Remark 4, page 15), which ensures that there exists at least one sampling design on Q with mean µ. The Lagrangian function is ( ) ) ( p(s) L(p(s), β, γ) = p(s) log p(s) − 1 − γ sp(s) − µ , −β pr (s) s∈Q
s∈Q
s∈Q
where β ∈ R and γ ∈ RN are the Lagrange multipliers. By setting the derivatives of L with respect to p(s) equal to zero, we obtain ∂L(p(s), β, γ) p(s) + 1 − β − γ s = 0, = log ∂p(s) pr (s)
5.2 General Exponential Designs
which gives
65
p(s) = pr (s) exp(γ s + β − 1).
By using the constraint (5.2), we get pr (s) exp γ s . s∈Q pr (s) exp γ s
p(s) =
Now, by replacing pr (s) by (5.1) we get the exponential designs ' ' 1 1 exp γ θn(s) s k∈U sk ! k∈U sk ! exp λ s ' ' pEXP (s, λ, Q) = = , 1 1 n(s) s exp γ exp λ θ s s∈Q k∈U sk ! s∈Q k∈U sk ! where λ = (λ1 · · · λN ) = (λ1 + log θ · · · λN + log θ) . The vector λ is identified by means of the constraint (5.3). We show that the problem of deriving λ from µ is one of the most intricate questions of exponential designs. 5.2.2 Exponential Designs (EXP) We refer to the following definition. Definition 45. A sampling design pEXP (.) on a support Q is said to be exponential if it can be written / 0 pEXP (s, λ, Q) = g(s) exp λ s − α(λ, Q) , where λ ∈ RN is the parameter, g(s) =
& 1 , sk !
k∈U
and α(λ, Q) is called the normalizing constant and is given by α(λ, Q) = log g(s) exp λ s. s∈Q
As it is the case for all of the exponential families, the expectation can be obtained by differentiating the normalizing constant sg(s) exp λ s ∂α(λ, Q) α (λ, Q) = = µ. = s∈Q ∂λ s∈Q g(s) exp λ s By differentiating the normalizing constant twice, we get the variance-covariance operator: α (λ, Q) ss g(s) exp λ s sg(s) exp λ s sg(s) exp λ s s∈Q s∈Q = − s∈Q s∈Q g(s) exp λ s s∈Q g(s) exp λ s s∈Q g(s) exp λ s = Σ.
66
5 Unequal Probability Exponential Designs
The characteristic function of an exponential design is given by / 0 g(s) exp(it + λ )s − α(λ, Q) = exp [α(it + λ, Q) − α(λ, Q)] . φEXP (t) = s∈Q
Simple random sampling designs are a particular case of exponential designs. Indeed, we have pEXP (s, 1 log θ, Q) = pSIMPLE (s, θ, Q), θ ∈ R+ , where 1 is a vector of N ones. The parameter λ can be split into two parts: λa , that is, the orthogonal → − projection of λ onto Q (the direction of Q, see Definition 9, page 12) and λb , that is, the orthogonal projection of λ onto Invariant Q (see Definition 11, page 12). We thus have that λ = λa + λb , and λa λb = 0. Moreover, we have the following result: → − Result 18. Let λa and λb be, respectively, the orthogonal projection on Q and Invariant Q. Then pEXP (s, λ, Q) = pEXP (s, λa + λb , Q) = pEXP (s, λa , Q) = pEXP (s, λa + bλb , Q), for any b ∈ R. Proof. pEXP (s, λa + bλb , Q) g(s)(exp λa s)(exp bλb s) g(s) exp(λa + bλb ) s = = . s∈Q g(s) exp(λa + bλb ) s s∈Q g(s)(exp λa s)(exp bλb s) Because λb ∈ Invariant (Q), (s−µ) λb = 0, and thus λb s = c, where c = µ λb for all s ∈ Q. Thus, pEXP (s, λa + bλb , Q) g(s)(exp λa s)(exp bc) g(s)(exp λa s) = = = pEXP (s, λa , Q). 2 s∈Q g(s)(exp λa s)(exp bc) s∈Q g(s)(exp λa s)
Example 8. If the support is Rn = s ∈ NN k∈U sk = n , then Invariant Rn = {u ∈ RN |u = a1 for all a ∈ R}, −→ N Rn = u ∈ R xk = 0 , k∈U
1 ¯ λb = 1 λk = 1λ, N k∈U
and ¯ · · · λk − λ ¯ · · · λN − λ), ¯ λa = (λ1 − λ where 1 is a vector of ones of RN .
5.3 Poisson Sampling Design With Replacement (POISSWR)
67
The main result of the theory of exponential families establishes that the mapping between the parameter λ and the expectation µ is bijective. Theorem 5. Let pEXP (s, λ, Q) be an exponential design on support Q. Then spEXP (s, λ, Q) µ(λ) = s∈Q ◦ → − is a homeomorphism of Q (the direction of Q) and Q (the interior of Q); that ◦ → − is, µ(λ) is a continuous and bijective function from Q to Q whose inverse is also continuous.
Theorem 5 is an application to exponential designs of a well-known result of exponential families. The proof is given, for instance, in Brown (1986, p. 74). Example 9. If the support is Rn , then µ(λ) is bijective from −→ N Rn = u ∈ R uk = 0 , k∈U
to ◦
Rn =
u ∈]0, n[ uk = n .
N
k∈U
Another important property of the exponential design is that the parameter does not change while conditioning with respect to a subset of the support. Result 19. Let Q1 and Q2 be two supports such that Q2 ⊂ Q1 . Also, let / 0 p1 (s) = g(s) exp λ s − α(λ, Q1 ) be an exponential design with support Q1 . Then p2 (s) = p1 (s|Q2 ) is also an exponential sampling design with the same parameter. The proof is obvious.
5.3 Poisson Sampling Design With Replacement (POISSWR) 5.3.1 Sampling Design Definition 46. An exponential design defined on R is called a Poisson sampling with replacement (POISSWR).
68
5 Unequal Probability Exponential Designs
The Poisson sampling with replacement can be derived from Definition 45, page 45. ) ( & 1 / 0 exp λ s − α(λ, R) , pPOISSWR (s, λ) = pEXP (s, λ, R) = sk ! k∈U
for all s ∈ R, and with λ ∈ RN . The normalizing constant is α(λ, R) = log
exp λ s & (exp λk )sk ' = log sk ! k∈U sk !
s∈R
s∈R k∈U
∞ & & (exp λk )j = log exp(exp λk ) = exp λk . = log j! j=1 k∈U
k∈U
Thus, µk = Σkk = Σk =
∂2
∂2
∂
k∈U
exp λk
∂λk
k∈U exp λk ∂λ2k
k∈U
= exp λk ,
= exp λk = µk ,
exp λ2k = 0, ∂λk ∂λ k∈U
with k = ,
and Σ = diag(µ1 · · · µk · · · µN ), which allows reformulating the definition of Poisson sampling with replacement. Definition 47. A sampling design pPOISSWR (., λ) on R is said to be a Poisson sampling with replacement if it can be written pPOISSWR (s, λ) =
& µsk e−µk k , sk !
k∈U
for all s ∈ R, and with µk ∈ R∗+ . The Sk ’s thus have independent Poisson distributions: Sk ∼ P(µk ). If the µk ’s are given, then λ can be computed by λk = log µk , k ∈ U. The characteristic function is φPOISSWR (t) = exp [α(it + λ, Q) − α(λ, Q)] = exp exp(itk + λk ) − exp λk = exp µk (exp itk − 1). k∈U
k∈U
k∈U
5.3 Poisson Sampling Design With Replacement (POISSWR)
69
5.3.2 Estimation The Hansen-Hurwitz estimator is Y!HH =
y k Sk . µk
k∈U
Its variance is var(Y!HH ) =
y2 k µk
k∈U
and can be estimated by var % 1 (Y!HH ) =
Sk
k∈U
yk2 . µ2k
However, because E [Sk |r(S)] =
µk if Sk > 0 1 − e−µk 0 if Sk = 0,
the improved Hansen-Hurwitz estimator can be computed Y!IHH =
yk E[Sk |r(S)] yk r(Sk ) = . µk 1 − e−µk
k∈U
Because and
k∈U
E[r(Sk )] = 1 − Pr(Sk = 0) = 1 − e−µk , var[r(Sk )] = E[r(Sk )] {1 − E[r(Sk )]} = e−µk 1 − e−µk ,
we obtain var(Y!IHH ) =
e−µk 1 − e−µk
k∈U
y2 yk2 k = , −µ 2 k (1 − e ) eµk − 1 k∈U
and var( % Y!IHH ) =
k∈U
(eµk
yk2 . − 1)(1 − e−µk )
The improvement brings an important decrease of the variance with respect to the Hansen-Hurwitz estimator. Finally, the Horvitz-Thompson estimator is Y!HT =
yk r(Sk ) yk r(Sk ) = E[r(Sk )] 1 − e−µk
k∈U
k∈U
and is equal to the improved Hansen-Hurwitz estimator.
70
5 Unequal Probability Exponential Designs
Algorithm 5.1 Sequential procedure for POISSWR Definition k : Integer; For k = 1, . . . , N do randomly select sk times unit k according to the Poisson distribution P(µk ); EndFor.
5.3.3 Sequential Procedure for POISSWR The standard sequential procedure (see Algorithm 3.2, page 34) applied to POISSWR gives the strictly sequential Algorithm 5.1.
5.4 Multinomial Design (MULTI) 5.4.1 Sampling Design Definition 48. An exponential design defined on Rn is called a multinomial design. The multinomial design can be derived from Definition 45, page 45. ) ( & 1 / 0 exp λ s − α(λ, Rn ) , pMULTI (s, λ, n) = pEXP (s, λ, Rn ) = sk ! k∈U
for all s ∈ Rn , and with λ ∈ RN . The normalizing constant is α(λ, Rn ) = log
s∈Rn
exp λ s ' k∈U sk !
& (exp λk )sk = log = log sk !
k∈U
s∈Rn k∈U
Thus, µk =
exp λk n!
n .
∂α(λ, Rn ) n exp λk = , ∂λk ∈U exp λ
and ⎧ 2 ∂ α(λ, Rn ) µk (n − µk ) ⎪ ⎪ = , ⎪ ⎨ ∂λ2 n Σk =
⎪ ⎪ ⎪ ⎩
k=
k
n exp λk exp λ µk µ ∂ α(λ, Rn ) = − , 2 = − n ∂λk ∂λ exp λ j j∈U 2
k = .
5.4 Multinomial Design (MULTI)
The variance-covariance operator is thus ⎛ µ (n−µ ) 1 1 · · · −µn1 µk n ⎜ .. .. .. ⎜ . . . ⎜ ⎜ k) Σ = var(S) = ⎜ −µn1 µk · · · µk (n−µ n ⎜ .. .. ⎜ ⎝ . . −µ1 µN −µk µN · · · n n
···
−µ1 µN n
··· .. . ···
−µk µN n
.. . .. .
µN (n−µN ) n
Moreover, we have ⎧ µk (n − µk ) µ2 (n − 1) ⎪ ⎪ + µ2k = µk + k , ⎨ n n µk = ⎪ ⎪ ⎩ − µk µ + µk µ = µk µ (n − 1) , n n and thus
⎧ ⎪ ⎪ ⎨
(n − µk ) , Σk n + µk (n − 1) = ⎪ 1 µk ⎪ ⎩− , n−1
71
⎞ ⎟ ⎟ ⎟ ⎟ ⎟. ⎟ ⎟ ⎠
k= k = ,
k= (5.4) k = .
The computation of µ allows reformulating the multinomial design in function of the µk ’s. Definition 49. A sampling design pMULTI (., λ, n) on Rn is said to be a multinomial design (MULTI) if it can be written pMULTI (s, λ, n) =
& n! (µk /n)sk , s1 ! . . . sk ! . . . sN ! k∈U
for all s ∈ Rn , and where µk ∈
R∗+
and
µk = n.
k∈U
This new formulation clearly shows that S has a multinomial distribution. The Sk ’s thus have non-independent binomial distributions: Sk ∼ B(n, µk /n). The invariant subspace spanned by support Rn is c × 1, for any value of c ∈ R. If the µk are fixed then λ can be computed by λk = log µk + c, k ∈ U, for any value of c ∈ R. The characteristic function is )n ( 1 φMULTI (t) = µk exp itk . n k∈U
72
5 Unequal Probability Exponential Designs
5.4.2 Estimation 1. The Hansen-Hurwitz estimator is Y!HH =
Sk
k∈U
and its variance is y2 1 k var(Y!HH ) = − µk n k∈U
(
)2 yk
k∈U
yk , µk 2 µk nyk = −Y . n2 µk k∈U
2. With Expression (5.4), the estimator (2.19), page 28, becomes )2 ( !HH y n Y k Sk − . var % 2 (Y!HH ) = (n − 1) µk n
(5.5)
k∈U
Using Result 9, page 27, estimator (2.18), page 28, becomes Sk y 2 S Σk k µ2k µk k∈U ∈U )2 ( yk n Y!HH = Sk − (n − 1) µk n k∈U " 2 y n n 2 k S . + − S k k µ2k n − 1 [n + µk (n − 1)]
% 2 (Y!HH ) + var % 1 (Y!HH ) = var
(5.6)
k∈U
As for SRSWR, this estimator (5.6) should never be used. 3. Because µk n πk = Pr(Sk > 0) = 1 − Pr(Sk = 0) = 1 − 1 − , n and πk = Pr(Sk > 0 and S > 0) = 1 − Pr(Sk = 0 or S = 0) = 1 − Pr(Sk = 0) − P (S = 0) + Pr(Sk = 0 and S = 0) µ n µk µk n µk n − 1− 1− +1− 1− = 1− 1− 1− − n n n n µk n µ n µk µk n = 1− + 1− − 1− , − n n n n the Horvitz-Thompson estimator is r(Sk )yk n . Y!HT = 1 − 1 − µnk k∈U Its variance and variance estimator can be constructed by means of the joint inclusion probabilities.
5.4 Multinomial Design (MULTI)
73
4. The construction of the improved Hansen-Hurwitz estimator seems to be an unsolved problem. The main difficulty consists of calculating E[Sk |r(S)]. A solution can be given for n = 3 but seems to be complicated for larger sample sizes. Case where n(S) = 3 Let ak = (0 · · · 0 *+,1 0 · · · 1) ∈ RN . Defining qk = µk /n, we have kth Pr [r(S) = ak ] = Pr [S = 3ak ] = qk3 ,
k ∈ U,
Pr [r(S) = ak + a ] = Pr(S = 2ak + a ) + Pr(S = ak + 2a ) = 3qk2 q + 3qk q2 = 3qk q (qk + q ), q = ∈ U, Pr [r(S) = ak + a + am ] = Pr(S = ak + a + am ) 3! = qk q qm = 6qk q qm . 1!1!1! The distribution of n[r(s)] can thus be computed Pr {n[r(S)] = 1} = qk3 , k ∈ U, ∈U
Pr {n[r(S)] = 2} =
3qk q (qk + q )
k∈U ∈U 2, and
1 if j = n 0 if j = n. / 0 N From λ, it is thus possible to compute α λ, Sz (UN recursive −k/) by means of 0 N ) by means Expression (5.29). Next qkj can be computed from α λ, Sz (UN −k of Expression (5.31). Finally, π can be derived from qkj , which allows defining an alternate procedure of deriving π from λ in place of Expression (5.13), page 81. Pr (r1 = j) =
5.6.11 Draw by Draw Procedure for CPS In order to have a fast implementation of the standard procedure, we use the following result from Chen et al. (1994). Result 29. Let S be a sample selected according to a CPS, A a subset of U and 1 /A E [Sk |Si = 1 for all i ∈ A] k ∈ qk (A) = n − card A 0 k ∈ A. Then qk (∅) = πk /n, and ⎧ qk (A) ⎪ ⎪ 1 qj (A) exp λj − exp λk ⎪ ⎪ k∈ / A, qk (A) = qj (A) ⎪ ⎨ n − cardA − 1 exp λj − exp λk qk (A ∪ {j}) = 1 − k∈A,q qk (A ∪ {j}) / k (A)=qj (A) ⎪ k∈ / A, qk (A) = qj (A) ⎪ ⎪ ⎪ card{k ∈ / A, qk (A) = qj (A)} ⎪ ⎩ 0 k ∈ A ∪ {j}. (5.32)
94
5 Unequal Probability Exponential Designs
Proof. Let a = cardA. If λj = λk , we have, for all j = k ∈ U \A, (exp λj ) sk exp λ s − (exp λk ) sj exp λ s s∈Sn−a (U \A)
s∈Sn−a (U \A)
⎛
⎜ = (exp λj ) ⎜ ⎝
s∈Sn−a (U \A) sk =1,sj =0
s∈Sn−a (U \A) sk =1,sj =1
⎛
⎜ − (exp λk ) ⎜ ⎝
exp λ s +
s∈Sn−a (U \A) sk =0,sj =j
= (exp λj − exp λk )
exp λ s +
⎞
⎟ exp λ s⎟ ⎠
s∈Sn−a (U \A) sk =1,sj =1
⎞ ⎟ exp λ s⎟ ⎠
sk sj exp λ s.
(5.33)
s∈Sn−a (U \A)
Moreover, because 1 s∈S (U \A) sk exp λ s n−a , k∈ / A, qk (A) = n−a s∈Sn−a (U \A) exp λ s qk (A) s∈S (U \A) sk exp λ s = n−a / A, , k ∈ qj (A) s∈Sn−a (U \A) sj exp λ s
and
we have qk (A) qj (A)
exp λj − exp λk
exp λj − exp λk (exp λj ) s∈Sn−a (U \A) sk exp λ s − (exp λk ) s∈Sn−a (U \A) sj exp λ s = . (exp λj − exp λk ) s∈Sn−a (U \A) sj exp λ s By (5.33), we obtain qk (A) qj (A)
exp λj − exp λk
exp λj − exp λk s∈Sn−a (U \A) sk sj exp λ s s∈Sn−a−1 [U \(A∪{j})] sk exp λ s = = s∈Sn−a (U \A) sj exp λ s s∈Sn−a−1 [U \(A∪{j})] exp λ s = (n − a − 1)qk (A ∪ {j}).
2
The standard draw by draw procedure given in Expression (3.3), page 35, gives Algorithm 5.9. Other procedures that implement the CPS design are proposed in Chen and Liu (1997).
5.8 Links Between Exponential Designs and Simple Designs
95
Algorithm 5.9 Draw by draw procedure for CPS 1. Compute λ from π by means of the Newton method as written in Expression (5.16). 2. Define A0 = ∅. 3. For j = 0, . . . , n − 1, do Select a unit from U with probability qk (Aj ) as defined in Expression (5.32); If kj is the selected unit, then Aj+1 = Aj ∪ {kj }; EndIf; EndFor. 4. The selected units are in the set An .
5.7 Links Between the Exponential Designs The following relations can be proved using the definition of conditioning with respect to a sampling design. • • • • •
pPOISSWR (s, λ|Rn ) = pMULTI (s, λ, n), pPOISSWR (s, λ|S) = pPOISSWOR (s, λ), pPOISSWR (s, λ|Sn ) = pCPS (s, λ, n), pMULTI (s, λ, n|Sn ) = pCPS (s, λ, n), pPOISSWOR (s, λ|Sn ) = pCPS (s, λ, n).
These relations are summarized in Figure 5.1 and Table 5.7, page 97. reduction
POISSWOR
conditioning on S POISSWR
conditioning on Sn conditioning on Sn
conditioning on Rn with 0 ≤ n ≤ N
CPS conditioning on Sn
MULTI
Fig. 5.1. Links between the main exponential sampling designs
5.8 Links Between Exponential Designs and Simple Designs The following relations are obvious: • pPOISSWR (s, 1 log θ) = pEPPOISSWR (s, θ), • pMULTI (s, 1 log θ, n) = pSRSWR (s, n),
96
5 Unequal Probability Exponential Designs
θ • pPOISSWOR (s, 1 log θ) = pBERN s, π = 1+θ • pCPS (s, 1 log θ, n) = pSRSWOR (s, n).
,
5.9 Exponential Procedures in Brewer and Hanif Brewer and Hanif (1983) have defined a set of properties in order to evaluate the sampling methods. These properties are listed in Table 5.6. Seven exponential procedures are described in Brewer and Hanif (1983); they are presented in Table 5.8, page 98. Most of them are due to H´ ajek. Because the way of deriving the inclusion probabilities from the parameter was not yet defined, H´ ajek proposed different working probabilities to apply the rejective Poisson sampling design. The H´ ajek procedures do not have probabilities of inclusion strictly proportional to size (not strpps). The Yates and Grundy (1953) procedure and the Carroll and Hartley (1964) methods are very intricate and can be used only for small sample sizes. Table 5.6. Abbreviations of Brewer and Hanif (1983) and definition of “exact” and “exponential” strpps strwor n fixed syst d by d rej ws ord unord inexact
Inclusion probability strictly proportional to size Strictly without replacement Number of units in sample fixed Systematic Draw by draw Rejective Whole sample Ordered Unordered Fails to satisfy at least one of the three properties: strpps, strwor, n fixed n=2 only Applicable for sample size equal to 2 only b est var Estimator of variance generally biased j p enum Calculation of joint inclusion probability in sample involves enumeration of all possible selections, or at least a large number of them j p iter Calculation of joint inclusion probability in sample involves iteration on computer not gen app Not generally applicable nonrotg Nonrotating exact Satisfies conjointly the three properties: strpps, strwor, n fixed exponential Procedure that implements an exponential design
var(Y )
var( Y )
Variance
Variance est.
Expectation µk Inclusion probability πk Variance Σkk Joint expectation µk , k = Covariance Σk , k = Basic est. Y µk
k∈U
y2 k Sk µ2k
k∈U
µk 1 − e−µk µk µ2k 0 yk Sk µk k∈U y2 k
random
n(S)
Sample size
k∈U
with repl.
µk (eitk − 1)
WOR/WR
exp λk
Replacement
exp
k∈U
φ(t)
α(λ, Q)
Char. function
Norming constant
R
Q
Support
k∈U
µk 1 − (1 − µk /n)n µk (n − µk )/n µk µ (n − 1)/n −µk µ /n yk Sk µk k∈U 2 nyk −Y µk k∈U
2 n YHH yk Sk − n−1 µk n
fixed
n Rn 1 log exp λk n! k∈U
n 1 µk exp itk n k∈U with repl.
k∈U
s n! µkk nn sk !
µsk e−µk k sk !
p(s)
Design k∈U
MULTI
POISSWR
Notation
k∈U
k∈U
s
(1 + exp λk )
S
πkk (1 − πk )1−sk
k∈U
Sk yk2
k∈U
1 − πk πk2
πk πk πk πk2 0 yk Sk πk k∈U 2 1 − πk yk πk
random
without repl.
{1 + πk (exp itk − 1)}
log
k∈U
POISSWOR
Table 5.7. Main exponential sampling designs
Sn
exp[λ s − α(λ, Sn )]
k∈U
not reducible
not reducible
πk (λ, Sn ) see Section 5.6.2 πk (λ, Sn ) see Section 5.6.4 πk (λ, Sn ) see Section 5.6.4 see Section 5.6.4 yk Sk πk
fixed
without repl.
not reducible
see Section 5.6.6
s∈Sn
CPS
31 40
30 40
29 40
28 40
27 39
14 31
5 24
Number Page
Table 5.8. Brewer and Hanif list of exponential procedures
Procedure name Yates-Grundy Rejective Procedure Principal reference: Yates and Grundy (1953) Other ref.: Durbin (1953), H´ ajek (1964) Carroll-Hartley Rejective Procedure Principal reference: Carroll and Hartley (1964) Other ref.: Rao (1963a), H´ ajek (1964), Rao and Bayless (1969), Bayless and Rao (1970), Cassel et al. (1993, p. 16) Poisson Sampling Principal reference: H´ ajek (1964) Other ref.: Ogus and Clark (1971), Brewer et al. (1972), Brewer et al. (1984), Cassel et al. (1993, p. 17) H´ ajek’s “Method I” Principal reference: H´ ajek (1964) H´ ajek’s “Method II” Principal reference: H´ ajek (1964) H´ ajek’s “Method III” Principal reference: H´ ajek (1964) H´ ajek’s “Method IV” Principal reference: H´ ajek (1964)
−++−−+−−++−−−−−−−+
−++−−+−−++−−−−−−−+
−++−−+−−++−−−−−−−+
−++−−+−−++−−−−−−−+
++−−+−−−++−−−−−−−+
+++−−+−−+−−−−+−−++
+−+−−+−−++−−−−−+−+
strpps strwor n fixed syst d by d rej ws ord unord inexact n=2 only b est var j p enum j p iter not gen app nonrotg exact exponential
6 The Splitting Method
6.1 Introduction The splitting method, proposed by Deville and Till´e (1998), is a general framework of sampling methods without replacement, with fixed sample size and unequal probabilities. The basic idea consists of splitting the inclusion probability into two or several new vectors. Next, one of these vectors is selected randomly in such a way that the average of the vectors is the vector of inclusion probabilities. This simple step is repeated until a sample is obtained. The splitting method is thus a martingale algorithm. It includes all the draw by draw and sequential procedures and it allows deriving a large number of unequal probability methods. Moreover, many well-known procedures of unequal probabilities can be formulated under the form of a splitting. The presentation can thus be standardized, which allows a simpler comparison of the procedures. In this chapter, we present the fundamental splitting algorithm. Two variants can be defined depending on whether a vector or a direction is chosen. Several applications are given: minimum support design, splitting into simple random sampling, pivotal procedure, and the generalized Sunter procedure. Next, a general method of splitting into several vectors is presented. The particular cases are the Brewer procedure, the eliminatory method, the Till´e elimination method, the generalized Midzuno procedure, and the Chao procedure.
6.2 Splitting into Two Vectors 6.2.1 A General Technique of Splitting into Two Vectors Consider a vector of inclusion probabilities π = (π1 · · · πN ) such that n(π) = πk = n, k∈U
100
6 The Splitting Method
and n is an integer. The general splitting technique into two vectors is presented in Algorithm 6.1. As shown in Figure 6.1, at each step t = 0, 1, 2, . . . , vector π(t) is split into two other vectors. The splitting method is based on a random transformation of vector π until a sample is obtained. Algorithm 6.1 General splitting method into two vectors 1. Initialize π(0) = π, and t = 0. 2. While π(t) ∈ / {0, 1}N do • Construct any pair of vectors π a (t), π b (t) ∈ RN , and a scalar α(t) ∈ [0, 1] such that b α(t)π a (t) + [1 − α(t)] π (t) = π(t), n [π a (t)] = n π b (t) = n [π(t)] , 0 ≤ πka (t) ≤ 1, and 0 ≤ πkb (t) ≤ 1;
a π (t) with probability α(t) • π(t + 1) = π b (t) with probability 1 − α(t); • t = t + 1; EndWhile. 3. The selected sample is S = π(t).
⎡
⎤ π1 (t) ⎢ .. ⎥ ⎢ . ⎥ ⎢ ⎥ ⎢ πk (t) ⎥ ⎢ ⎥ ⎢ . ⎥ ⎣ .. ⎦ πN (t)
HH HH 1 − α(t) α(t) H HH j H ⎡
⎤ π1a (t) ⎢ .. ⎥ ⎢ . ⎥ ⎢ a ⎥ ⎢ πk (t) ⎥ ⎢ ⎥ ⎢ . ⎥ ⎣ .. ⎦ a πN (t)
⎡
⎤ π1b (t) ⎢ . ⎥ ⎢ . ⎥ ⎢ . ⎥ ⎢ b ⎥ ⎢ πk (t) ⎥ ⎢ ⎥ . ⎢ . ⎥ ⎣ . ⎦ b πN (t)
Fig. 6.1. Splitting into two vectors
6.2 Splitting into Two Vectors
101
Vector π(t) is a martingale because E [π(t)|π(t − 1), . . . , π(1)] = π(t − 1). It follows that E [π(t)] = π, and thus the splitting method selects a sample with the inclusion probabilities equal to π. At each step of the algorithm, the problem is reduced to another sampling problem with unequal probabilities. If the splitting is such that one or several of the πka and the πkb are equal to 0 or 1, the sampling problem will be simpler at the next step. Indeed, once a component of π(t) becomes an integer, it must remain an integer for all the following steps of the method. The splitting method allows defining a large family of algorithms that includes many known procedures. 6.2.2 Splitting Based on the Choice of π a (t) A first way to construct π a (t), π b (t), and α(t) consists of arbitrarily fixing π a (t) and then of deriving appropriate values for π b (t) and α(t), which gives Algorithm 6.2. Algorithm 6.2 Splitting method based on the choice of π a (t) Initialize π(0) = π; For t = 0, 1, 2 . . . , and until a sample is obtained, do 1. Generate any vector of probabilities π a (t) such that n[π a (t)] = n(π) and πka (t) = πk (t) if πk (t) ∈ {0, 1} . The vector π a (t) can be generated randomly or not; 2. Define α(t) as the largest real number in [0, 1] that satisfies 0≤ that is,
πk (t) − α(t)πka (t) ≤ 1, for all k ∈ U ; 1 − α(t) ⎡
α(t) = min ⎣
min
k∈U 0